Systems and methods for aggregating related inputs using finite-state devices and extracting meaning from multimodal inputs using aggregation
When using finite-state methods to represent inputs, it is beneficial to combine related inputs and to represent combinations of related inputs in the finite-state device. In multimodal recognition systems and methods that integrate different recognized modes, combining related inputs helps the unification process.
Latest AT&T Patents:
- FORWARD COMPATIBLE NEW RADIO SIDELINK SLOT FORMAT SIGNALLING
- HOMOGLYPH ATTACK DETECTION
- METHODS, SYSTEMS, AND DEVICES FOR MASKING CONTENT TO OBFUSCATE AN IDENTITY OF A USER OF A MOBILE DEVICE
- CUSTOMIZABLE AND LOW-LATENCY ARCHITECTURE FOR CELLULAR CORE NETWORKS
- LOCATION AWARE ASSIGNMENT OF RESOURCES FOR PUSH TO TRANSFER (PTT) COMMUNICATION SYSTEMS IN A FIFTH GENERATION (5G) NETWORK OR OTHER NEXT GENERATION WIRELESS COMMUNICATION SYSTEM
[0001] This non-provisional application claims the benefit of U.S. provisional application No. 60/313,121, filed on Aug. 15, 2001, which is incorporated herein by reference in its entirety.
BACKGROUND OF THE INVENTION[0002] 1. Field of Invention
[0003] This invention is directed to aggregating related inputs when using finite-state methods.
[0004] 2. Description of Related Art
[0005] Multimodal interfaces allow input and/or output to be conveyed over multiple different channels, such as speech, graphics, gesture and the like. Multimodal interfaces enable more natural and effective interaction, because particular modes are best-suited for particular kinds of content. Multimodal interfaces are likely to play a critical role in the ongoing migration of interaction from desktop computing to wireless portable computing devices, such as personal digital assistants, like the Palm Pilot®, digital cellular telephones, public information kiosks that are wirelessly connected to the Internet or other distributed networks, and the like. One barrier to adopting such wireless portable computing devices is that they offer limited screen real estate, and often have limited keyboard interfaces, if any keyboard interface at all.
[0006] To realize the full potential of such wireless portable computing devices, multimodal interfaces need to support not just input from multiple modes. Rather, multimodal interfaces also need to support synergistic multimodal utterances that are optimally distributed over the various available modes. In order to achieve this, the content from different modes needs to be effectively integrated.
[0007] One previous attempt at integrating the content from the different modes is disclosed in “Unification-Based Multimodal Integration”, M. Johnston et al., Proceedings of the 35th ACL, Madrid Spain, p. 281-288, 1997 (Johnston 1), incorporated herein by reference in its entirety. Johnston 1 disclosed a pen-based device that allows a variety of gesture utterances to be input through a gesture mode, while a variety of speech utterances can be input through a speech mode.
[0008] In Johnston 1, a unification operation over typed feature structures was used to model the integration between the gesture mode and the speech mode. Unification operations determine the consistency of two pieces of partial information. If the two pieces of partial information are determined to be consistent, the unification operation combines the two pieces of partial information into a single result. Unification operations were used to determine whether a given piece of gestural input received over the gesture mode was compatible with a given piece of spoken input received over the speech mode. If the gestural input was determined to be compatible with the spoken input, the two inputs were combined into a single result that could be further interpreted.
[0009] In Johnston 1, typed feature structures were used as a common meaning representation for both the gestural inputs and the spoken inputs. In Johnston 1, the multimodal integration was modeled as a cross-product unification of feature structures assigned to the speech and gestural inputs. While the technique disclosed in Johnston 1 overcomes many of the limitations of earlier multimodal systems, this technique does not scale well to support multi-gesture utterances, complex unimodal gestures, or other modes and combinations of modes. To address these limitations, the unification-based multimodal integration technique disclosed in Johnston 1 was extended in “Unification-Based Multimodal Parsing”, M. Johnston, Proceedings of COLING-ACL 98, p. 624-630, 1998 (Johnston 2), herein incorporated by reference in its entirety. The multimodal integration technique disclosed in Johnston 2 uses a multi-dimensional chart parser. In Johnston 2, elements of the multimodal input are treated as terminal edges by the parser. The multimodal input elements are combined together in accordance with a unification-based multimodal grammar. The unification-based multimodal parsing technique disclosed in Johnston 2 was further extended in “Multimodal Language Processing”, M. Johnston, Proceedings of ICSLP 1998, 1998 (published on CD-ROM only) (Johnston 3), incorporated herein by reference in its entirety.
[0010] Johnston 2 and 3 disclosed how techniques from natural language processing can be adapted to support parsing and interpretation of utterances distributed over multiple modes. In the approach disclosed by Johnston 2 and 3, speech and gesture recognition produce n-best lists of recognition results. The n-best recognition results are assigned typed feature structure representations by speech interpretation and gesture interpretation components. The n-best lists of feature structures from the spoken inputs and the gestural inputs are passed to a multi-dimensional chart parser that uses a multimodal unification-based grammar to combine the representations assigned to the input elements. Possible multimodal interpretations are then ranked. The optimal interpretation is then passed on for execution.
[0011] Further, in Johnston 1-3, gestural inputs are assigned typed feature structures by the gesture representation agents. Using feature structures as a semantic representation framework allow for the specification of partial meanings. Spoken or gestural input which partially specify a command can be represented as an underspecified feature structure in which certain features are not instantiated. Adopting typed feature structures facilitates the statement of constraints on integration. For example, if a given speech input can be integrated with a line gesture, it can be assigned a feature structure with an underspecified location feature whose value is required to be of type line.
SUMMARY OF THE INVENTION[0012] However, the unification-based approach disclosed in Johnston 1-Johnston 3 does not allow for tight coupling of multimodal parsing with speech and gesture recognition. Compensation effects are dependent on the correct answer appearing in each of the n-best list of interpretations obtained from the recognitions obtained from the inputs of each mode. Moreover, multimodal parsing cannot directly influence the progress of either speech recognition or gesture recognition. The multi-dimensional parsing approach is also subject to significant concerns in terms of computational complexity. In the worst case, for the multi-dimensional parsing technique disclosed in Johnston 2, the number of parses to be considered is exponential relative to the number of input elements and the number of interpretations the input elements have. This complexity is manageable when the inputs yield only n-best results for small n. However, the complexity quickly gets out of hand if the inputs are sizable lattices with associated probabilities.
[0013] The unification-based approach also runs into significant problems when choosing between multiple competing parses and interpretations. Probabilities associated with composing speech events and multiple gestures need to be combined. Uni-modal interpretations need to be compared to multimodal interpretations and so on. While this can all be achieved using the unification-based approach disclosed in Johnston 1-Johnston 3, significant post-processing of sets of competing multimodal interpretations generated by the multimodal parser will be involved.
[0014] An alternative to the unification-based multimodal parsing technique disclosed in Johnston 3 is discussed in “Finite-state Multimodal Parsing and Understanding”, M. Johnston and S. Bangalore, Proceedings of COLING 2000, Saarbrucken, Germany, 2000, (Johnston 4) and in U.S. patent application Ser. No. 09/904,253, each incorporated herein by reference in its entirety. In Johnston 4, multimodal parsing, understanding and/or integration are achieved using a weighted finite-state device which takes speech and gesture streams as inputs and outputs their joint interpretation. This finite state approach is significantly more efficient, enables tight-coupling of multimodal understanding with speech recognition, and provides a general probabilistic framework for multimodal ambiguity resolution.
[0015] In Johnston 4 and the incorporated 253 application, language and gesture input streams are parsed and integrated by a single weighted finite-state device. This single weighted finite-state device provides language models for speech and gesture recognition and composes the meaning content from the speech and gesture input streams into a single semantic representation. Thus, Johnston 4 and the incorporated 253 application not only address multimodal language recognition, but also encode the semantics as well as the syntax into a single weighted finite-state device. Compared to the previous approaches for integrating multimodal input streams, such as those described in Johnston 1-3, which compose elements from n-best lists of recognition results, Johnston 4 and the incorporated 253 application provide the potential for direct compensation among the various multimodal input modes.
[0016] Further, in Johnston 4 and the incorporated 253 application, the structure interpretation of the multimodal commands is captured in a multimodal context-free grammar in which the multimodal aspects of the context free grammar are apparent in the terminals. The terminals contain three components, W, G and M, where W is the spoken language stream, G is the gesture stream, and M is the combined meaning. The non-terminals in the multimodal grammar are atomic symbols. In some exemplary embodiments of Johnston 4 and the incorporated 253 application, the gesture symbols G can be organized into a type hierarchy reflecting the ontology of the entities in the application domain. For example, a pointing gesture may be assigned the general semantic type “G”. This general semantic gesture “G” may have various subtypes, such as “Go” and “Gp”. In various exemplary embodiments, “Go” represents a gesture made against an organization object. In various exemplary embodiments, “Gp” represents a gesture made against a person object. Furthermore, the “Gp” type gesture may itself have subtypes, such as, for example, “Gpm” and “Gpf” for objects that respectively represent male and female persons. Compared with a feature-based multimodal grammar, these semantic types constitute a set of atomic categories which make the relevant distinctions for gesture events to predict speech events and vice versa.
[0017] Johnston 4 and the incorporated 253 application allow the gestural input to dynamically alter the language model used for speech recognition. In addition, the approach provided in Johnston 4 and the incorporated 253 application reduce the computational complexity of multi-dimensional multimodal parsing. In particular, the weighted finite-state devices, used in Johnston 4 and the incorporated 253 application, provide a well-understood probabilistic framework for combining the probability distributions associated with the speech and gesture or other input modes and for selecting among multiple competing multimodal interpretations.
[0018] Systems and methods for recognizing and representing gestural input have been very limited. Both the feature-based multimodal grammar provided in Johnston 1-3 and the multimodal context free grammar disclosed in Johnston 4 and the incorporated 253 application may be used to capture the structure and interpretation of multimodal commands. However, neither approach allows for efficiently and generically representing arbitrary gestures.
[0019] In addition, in the unification-based multimodal grammar disclosed in Johnston 1-3, spoken phrases and physical gestures are assigned typed feature structures by the natural language and gesture interpretation systems, respectively. In particular, each gesture feature structure includes a content portion that allows the specific location of the gesture on the gesture input portion of the multimodal user input device 400 to be specified. Thus, in the unification-based approach to multimodal integration disclosed in Johnston 1-3, specific content of an input gesture is captured by feature sharing.
[0020] However, in order to capture multimodal integration using finite-state methods, it is necessary to abstract over certain aspects of the gestural content. When using finite-state automata, a unique identifier is needed for each object or location in the gesture input portion 410 that a user can gesture on. For example, in the exemplary embodiment shown in FIG. 5, the finite-state automata would need to include a unique identifier for each object 412. In particular, as part of the composition process performed by the gesture recognition system 200, as well as the various composition processes described below, these identifiers would need to be copied from the gesture feature lattice 220 into the semantic representation that is represented by the gesture recognition lattice 255, which is generated by the gesture meaning subsystem 250. However, it is generally difficult, if not effectively impossible to capture all of different possible sequences of coordinates that occur in a gestural input in order to copy them from the gesture input tape to the meaning output tape of the finite-state automaton.
[0021] In the finite-state approach used in the systems and methods according to this invention, one possible, but ultimately unworkable solution, would be to incorporate all of the different possible identifiers for all of the different possible elements of the gesture input device 410, against which a gesture could be made, into the finite-state automata. For example, for an object having an identifier “object identifier 345”, an arc in the lattice would need to be labeled with that identifier in order to transfer that piece of information from the gesture tape to the meaning tape of the finite-state automaton. Moreover, the arc for each different identifier would have to be repeated numerous times in the network wherever this transfer of information would be needed. Furthermore, the various arcs would have to be updated as the underlying objects within the gesture input portion 410 were updated or changed.
[0022] Various exemplary embodiments of the systems and methods disclosed in Johnston 4 and the incorporated 253 application overcome this problem by storing the specific identifiers of the elements within the gesture input portion 410 associated with incoming gestures into a finite set of variables labeled “e1”, “e2”, “e3”, . . . . Then, in place of the specific object identifier, the labels of the variables storing that specific object identifier are incorporated into the various finite-state automata. Thus, instead of having arcs labeled with specific values in the finite-state automata, the finite-state automata include arcs labeled with the labels of the variables.
[0023] In Johnston 4 and the incorporated 253 application, therefore, instead of having the specific values for each possible object identifier in a finite-state automaton, that finite-state automaton instead incorporates the transitions “&egr;e1:e1”, “&egr;:e2:e2”, “&egr;:e3:e3”, . . . in each location in the finite-state automaton where specific content needs to be transferred from the gesture tape to the meaning tape.
[0024] For example, assuming a user using the gesture input portion 410 shown in FIG. 5 made the gestures 414 shown in FIG. 5, the gesture recognition system 200 would output, as the gesture recognition lattice 255, the finite-state machine shown in FIG. 8. In this case, as shown in FIG. 8, the arc labeled “Gp” corresponds to the gesture made against a person object while the arc labeled “Go” represents a gesture made against an organization object. The arc labeled “e1” stores the identifier of the person object 412,—in this case, the person object 412 labeled “Robert DeNiro”—against which the person-object gesture “Gp” 414 was made. Similarly, the arc labeled “e2” represents the variable storing the identifier of the organization object 412—in this case “Monumental Pictures”—against which the organization gesture 414 was made.
[0025] In Johnston 4 and the incorporated 253 application, during operation, the gesture recognition system 200 empties the variables e1, e2, e3 . . . after each multimodal command, so that all of the variables can be reused after each multimodal command. This allows the finite-state automaton to be built using a finite set of variables. However, this limits the number of distinct gesture events in a single utterance to no more than the available number of variables, because the number of gestural inputs that can be handled is limited by the number of variables used. Further, if a large number of variables are used, then similarly to using a large number of specific values, the resulting multimodal finite-state device increases significantly in size because an arc is needed for each variable in every place where gesture content can be copied onto the meaning tape.
[0026] In addition, when a single gesture is formed by two or more temporally and/or spatially related gesture components, the gesture components can be combined into a single gesture during the recognition process or by the multimodal parser/meaning recognition system. After the gesture features are extracted, the gesture recognition subsystem combines the corresponding, temporally-related and/or spatially-related gestures into a lattice of one or more recognized possible gesture sequences that represent how the recognized gestures follow each other in time and/or are positioned relative to each other in space.
[0027] The systems and methods disclosed in Johnston 4 and the incorporated 253 application, use certain simplifying assumptions with respect to temporal constraints. In multi-gesture utterances, the primary function of temporal constraints is to force an order on the gestures. For example, if a user generates the spoken utterance “move this here” and simultaneously makes two gestures, then the first gesture corresponds to the spoken utterance “this”, while the second gesture corresponds to the spoken utterance “here”. In the various exemplary embodiments of the systems and methods according to Johnston 4 and the incorporated 253 application, the multimodal grammars encode order, but do not impose explicit temporal constraints.
[0028] However, there are multimodal applications in which more specific temporal constraints are relevant. For example, specific temporal constraints can be relevant in selecting among unimodal and multimodal interpretations. That is, for example, if a gesture is temporally distant from the speech, then the unimodal interpretation should be treated as having a higher probability of being correct.
[0029] This invention provides systems and methods for representing the possible associations of a plurality of related inputs.
[0030] This invention separately provides systems and methods for representing the possible associations of a plurality of related inputs while maintaining a record of the original inputs.
[0031] This invention separately provides systems and methods for representing the possible associations of a plurality of temporally-related and/or spatially-related inputs.
[0032] This invention separately provides systems and methods for representing the possible associations of a plurality of temporally-related and/or spatially-related inputs while maintaining a record of the original inputs.
[0033] This invention separately provides systems and methods for representing possible interpretations of a plurality of related gestures.
[0034] This invention separately provides systems and methods for representing possible interpretations of a plurality of temporally-related and/or spatially-related gestures.
[0035] This invention separately provides systems and methods for aggregating a plurality of related inputs when using finite-state devices.
[0036] This invention separately provides systems and methods for aggregating a plurality of related gestures when using finite-state devices.
[0037] This invention separately provides systems and methods for aggregating a plurality of temporally-related and/or spatially-related inputs when using finite-state devices.
[0038] This invention separately provides systems and methods for aggregating a plurality of temporally and/or spatially related gestures when using finite-state devices.
[0039] This invention separately provides systems and methods for aggregating a plurality of temporally and/or spatially related gestures while maintaining an indication of the original gestures.
[0040] In various exemplary embodiments, the systems and methods according to this invention combine related inputs of a single mode of a multimodal recognition device by receiving a plurality of inputs of the single mode of an input portion of the multimodal device. An input lattice, representing the plurality of inputs is generated and at least two related inputs in the input lattice are combined. At least one branch is added to the input lattice to represent the combination of the at least two related inputs.
[0041] In various exemplary embodiments, the systems and methods according to this invention combine adjacent inputs of a multimodal device by receiving a plurality of inputs of one mode of the multimodal device in an input portion of the multimodal device. An input lattice representing the plurality of inputs is generated and at least two adjacent inputs on the input lattice are combined. At least one branch, representing the combination of related inputs, is added to the input lattice.
[0042] These and other features and advantages of this invention are described in, or are apparent from, the following detailed description of various exemplary embodiments of the systems and methods according to this invention.
BRIEF DESCRIPTION OF THE DRAWINGS[0043] Various exemplary embodiments of this invention will be described in detail, with reference to the following figures, wherein:
[0044] FIG. 1 is a block diagram illustrating one exemplary embodiment of a conventional automatic speech recognition system usable with a multimodal meaning recognition system according to this invention;
[0045] FIG. 2 is a block diagram illustrating a first exemplary embodiment of a multimodal user input device and one exemplary embodiment of a multimodal meaning recognition system according to this invention;
[0046] FIG. 3 is a block diagram illustrating in greater detail one exemplary embodiment of the gesture recognition system of FIG. 2;
[0047] FIG. 4 is a block diagram illustrating in greater detail one exemplary embodiment of the multimodal parser and meaning recognition system of FIG. 2;
[0048] FIG. 5 is a block diagram illustrating in greater detail the first exemplary embodiment of the multimodal user input device of FIG. 2;
[0049] FIG. 6 is one exemplary embodiment of a three-tape multimodal grammar fragment usable by the multimodal meaning recognition system according to this invention;
[0050] FIG. 7 is one exemplary embodiment of a multimodal three-tape finite-state automaton representing the multimodal context free grammar fragment shown in FIG. 6;
[0051] FIG. 8 is one exemplary embodiment of a gesture finite-state machine generated by recognizing the gesture inputs shown in the exemplary embodiment of the multimodal user input device shown in FIG. 5;
[0052] FIG. 9 is a second exemplary embodiment of a multimodal user input device;
[0053] FIG. 10 is an exemplary embodiment of an unimodal pen command inputted into the second exemplary embodiment of the multimodal user input device shown in FIG. 9;
[0054] FIG. 11 is one exemplary embodiment of two area gestures input into the multimodal user input device of FIG. 9;
[0055] FIG. 12 is one exemplary embodiment of phone query callouts resulting from a user's request input into the multimodal user input device of FIG. 9;
[0056] FIG. 13 is one exemplary embodiment of a gesture input finite-state transducer generated by recognizing the gesture inputs of a user inquiring about two restaurants;
[0057] FIG. 14 is one exemplary embodiment of a gesture input finite-state transducer generated by recognizing the gesture inputs of a user inquiring about a restaurant and a theatre;
[0058] FIG. 15 is one exemplary embodiment of a command requesting the multimodal user input device shown in FIG. 9 to compare a large set of restaurants;
[0059] FIG. 16 is one exemplary embodiment of a multimodal subway-route-information application displayed on the multimodal user input device shown in FIG. 9;
[0060] FIG. 17 is one exemplary embodiment of a resulting gesture lattice generated using one exemplary embodiment of a method for gesture representation according to this invention;
[0061] FIG. 18 is one exemplary embodiment of a three-tape multimodal finite-state automaton usable to recognize the multimodal inputs received from the exemplary embodiment of the multimodal user input device shown in FIG. 9;
[0062] FIG. 19 is one exemplary embodiment of a multimodal grammar fragment usable by the multimodal recognition system according to this invention;
[0063] FIG. 20 is one exemplary embodiment of a gesture-to-speech finite-state transducer that represents the relationship between speech and gesture for the exemplary embodiment of the multimodal user input device shown in FIG. 9;
[0064] FIG. 21 is one exemplary embodiment of a speech/gesture/meaning finite-state transducer that represents the relationship between the combined speech and gesture symbols and the semantic meaning of the multimodal input for the exemplary embodiment of the multimodal input device shown in FIG. 9;
[0065] FIG. 22 is one exemplary embodiment of a gesture finite-state transducer generated by recognizing the gesture inputs shown in the exemplary embodiment of the multimodal user input device shown in FIG. 9;
[0066] FIG. 23 is one exemplary embodiment of a gesture finite-state machine generated by recognizing the gesture inputs shown in the exemplary embodiment of the multimodal user input device shown in FIG. 9;
[0067] FIG. 24 is a flowchart outlining one exemplary embodiment of a method for extracting meaning from a plurality of multimodal inputs using abstraction;
[0068] FIG. 25 is one exemplary embodiment of a gesture/language finite-state transducer illustrating the composition of the gesture finite-state machine shown in FIG. 23 with the gesture-to-speech finite-state transducer shown in FIG. 20;
[0069] FIG. 26 is one exemplary embodiment of a speech input lattice generated by recognizing the speech input received by the exemplary embodiment of the multimodal user input device shown in FIG. 9;
[0070] FIG. 27 illustrates one exemplary embodiment of a gesture/speech finite-state transducer generated by composing the gesture/language finite-state transducer shown in FIG. 25 with the speech input lattice shown in FIG. 26;
[0071] FIG. 28 is one exemplary embodiment of a gesture/speech finite-state machine obtained from the gesture/speech finite-state transducer shown in FIG. 27;
[0072] FIG. 29 is one exemplary embodiment of a finite-state transducer, obtained from composing the gesture/speech finite-state machine shown in FIG. 27 with the speech/gesture/meaning finite-state transducer shown in FIG. 21, which extracts the meaning from the multimodal gestural and spoken inputs received when using the exemplary embodiment of the multimodal user input device shown in FIG. 9;
[0073] FIG. 30 is one exemplary embodiment of a finite-state machine generated by taking a projection of the finite-state transducer shown in FIG. 29;
[0074] FIG. 31 is one exemplary embodiment of a finite-state machine generated by composing the finite-state machine shown in FIG. 30 with the gesture finite-state transducer shown in FIG. 22; and
[0075] FIG. 32 is one exemplary illustration of a string of meaning read from the finite-state machine shown in FIG. 31.
DETAILED DESCRIPTION OF EXEMPLARY EMBODIMENTS[0076] FIG. 1 illustrates one exemplary embodiment of an automatic speech recognition system 100 usable with the multimodal recognition and/or meaning system 1000 according to this invention that is shown in FIG. 2. As shown in FIG. 1, automatic speech recognition can be viewed as a processing pipeline or cascade.
[0077] In each step of the processing cascade, one or two lattices are input and composed to produce an output lattice. In automatic speech recognition and in the following description of the exemplary embodiments of the systems and methods of this invention, the term “lattice” denotes a directed and labeled graph, which is possibly weighted. In each lattice, there is typically a designated start node “s” and a designated final node “t”. Each possible pathway through the lattice from the start node s to the final node t induces a hypothesis based on the arc labels between each pair of nodes in the path. For example, in a word lattice, the arc labels are words and the various paths between the start node s and the final node t form sentences. The weights on the arcs on each path between the start node s and the final node t are combined to represent the likelihood that that path will represent a particular portion of the utterance.
[0078] As shown in FIG. 1, one exemplary embodiment of a known automatic speech recognition system 100 includes a signal processing subsystem 110, an acoustic model lattice 120, a phonetic recognition subsystem 130, a lexicon lattice 140, a word recognition subsystem 150, a grammar or language model lattice 160, and a task recognition subsystem 170. In operation, uttered speech is input via a microphone, which converts the sound waves of the uttered speech into an electronic speech signal. The electronic speech signal is input to the signal processing subsystem 110 on a speech signal input line 105. The signal processing subsystem 110 digitizes the electronic speech signal to generate a feature vector lattice 115. The feature vector lattice 115 is a lattice of acoustic feature vectors. The feature vector lattice 115 is input along with the acoustic model lattice 120 to the phonetic recognition subsystem 130. The acoustic model lattice 120 represents a set of acoustic models and is applied to transform the feature vector lattice 115 into a phone lattice. Each node of the phone lattice represents a spoken sound, such as, for example, the vowel /e/ in “bed”.
[0079] The phone lattice 135 is input along with the lexicon lattice 140 into the word recognition subsystem 150. The lexicon lattice 140 describes different pronunciations of various words and transforms the phone lattice 135 into a word lattice 155. The word lattice 155 is then input, along with the grammar or language model lattice 160, into the utterance recognition subsystem 170. The grammar or language model lattice 160 represents task-specific information and is used to extract the most likely sequence of uttered words from the word lattice 155. Thus, the utterance recognition subsystem 170 uses the grammar or language model lattice 160 to extract the most likely sentence or other type of utterance from the word lattice 155. In general, the grammar or language model lattice 160 will be selected based on the task associated with the uttered speech. The most likely sequence of words, or the lattice of n most-likely sequences of words, is output as the recognized utterance 175.
[0080] In particular, one conventional method of implementing automatic speech recognition forms each of the acoustic model lattice 120, the lexicon lattice 140 and the grammar or language model lattice 160 as a finite-state transducer. Thus, each of the phonetic recognition subsystem 130, the word recognition subsystem 150, and the utterance recognition 170 performs a generalized composition operation between its input finite-state transducers. In addition, the signal processing subsystem 110 outputs the features vector lattice 115 as a finite-state transducer.
[0081] Conventionally, the grammar or language model lattice 160 is predetermined and incorporated into the automatic speech recognition system 100 based on the particular recognition task that the automatic speech recognition system 100 is to perform. In various exemplary embodiments, any of the acoustic model lattice 120, the lexicon lattice 140 and/or the grammar or language model 160 can be non-deterministic finite-state transducers. In this case, these non-deterministic finite-state transducers can be determinized using the various techniques disclosed in “Finite-state transducers in Language and Speech Processing”, M. Mohri, Computational Linguistics, 23:2, p. 269-312, 1997, U.S. patent application Ser. No. 09/165,423, filed Oct. 2, 1998, and/or U.S. Pat. No. 6,073,098 to Buchsbaum et al., each incorporated herein by reference in its entirety.
[0082] In contrast, in various exemplary embodiments of the systems and methods according to this invention, in the multimodal recognition or meaning system 1000 shown in FIG. 2, the automatic speech recognition system 100 uses a grammar or language model lattice 160 that is obtained from the recognized gestural input received in parallel with the speech signal 105. This is shown in greater detail in FIG. 2. In this way, the output of the gesture recognition system 200 can be used to compensate for uncertainties in the automatic speech recognition system.
[0083] Alternatively, in various exemplary embodiments of the systems and methods according this invention, the output of the automatic speech recognition system 100 and output of the gesture recognition system 200 can be combined only after each output is independently obtained. In this way, it becomes possible to extract meaning from the composition of two or more different input modes, such as the two different input modes of speech and gesture.
[0084] Furthermore, it should be appreciated that, in various exemplary embodiments of the systems and methods according to this invention, the output of the gesture recognition system 200 can be used to provide compensation to the automatic speech recognition system 100. Additionally, their combined output can be further processed to extract meaning from the combination of the two different input modes. In general, when there are two or more different input modes, any of one or more of the input modes can be used to provide compensation to one or more other ones of the input modes.
[0085] Thus, it should further be appreciated that, while the following detailed description focuses on speech and gesture as the two input modes, any two or more input modes that can provide compensation between the modes, which can be combined to allow meaning to be extracted from the two or more recognized outputs, or both, can be used in place of, or in addition to, the speech and gesture input modes discussed herein.
[0086] In particular, as shown in FIG. 2, when speech and gesture are the implemented input modes, a multimodal user input device 400 includes a gesture input portion 410 and a speech input portion 420. The gesture input portion 410 outputs a gesture signal 205 to a gesture recognition system 200 of the multimodal recognition and/or meaning system 1000. At the same time, the speech input portion 420 outputs the speech signal 105 to the automatic speech recognition system 100. The gesture recognition system 200 generates a gesture recognition lattice 255 based on the input gesture signal 205 and outputs the gesture recognition lattice 255 to a multimodal parser and meaning recognition system 300 of the multimodal recognition and/or meaning system 1000.
[0087] In those various exemplary embodiments that provide compensation between the gesture and speech recognition systems 200 and 100, the multimodal parser/meaning recognition system 300 generates a new grammar or language model lattice 160 for the utterance recognition subsystem 170 of the automatic speech recognition system 100 from the gesture recognition lattice 255. In particular, this new grammar or language model lattice 160 generated by the multimodal parser/meaning recognition system 300 is specific to the particular sets of gestural inputs generated by a user through the gesture input portion 410 of the multimodal user input device 400. Thus, this new grammar or language model lattice 160 represents all of the possible spoken strings that can successfully combine with the particular sequence of gestures input by the user through the gesture input portion 410. That is, the recognition performed by the automatic speech recognition system 100 can be improved because the particular grammar or language model lattice 160 being used to recognize that spoken utterance is highly specific to the particular sequence of gestures made by the user.
[0088] The automatic speech recognition system 100 then outputs the recognized possible word sequence lattice 175 back to the multimodal parser/meaning recognition system 300. In those various exemplary embodiments that do not extract meaning from the combination of the recognized gesture and the recognized speech, the recognized possible word sequences lattice 175 is then output to a downstream processing task. The multimodal recognition and/or meaning system 1000 then waits for the next set of inputs from the multimodal user input device 400.
[0089] In contrast, in those exemplary embodiments that additionally extract meaning from the combination of the recognized gesture and the recognized speech, the multimodal parser/meaning recognition system 300 extracts meaning from the combination of the gesture recognition lattice 255 and the recognized possible word sequences lattice 175. Because the spoken utterances input by the user through the speech input portion 420 are presumably closely related to the gestures input at the same time by the user through the gesture input portion 410, the meaning of those gestures can be tightly integrated with the meaning of the spoken input generated by the user through the speech input portion 420.
[0090] The multimodal parser/meaning recognition system 300 outputs a recognized possible meaning lattice 375 in addition to, or in place of, one or both of the gesture recognition lattice 255 and/or the recognized possible word sequences lattice 175. In various exemplary embodiments, the multimodal parser and meaning recognition system 300 combines the recognized lattice of possible word sequences 175 generated by the automatic speech recognition system 100 with the gesture recognition lattice 255 output by the gesture recognition system 200 to generate the lattice of possible meaning sequences 375 corresponding to the multimodal gesture and speech inputs received from the user through the multimodal user input device 400.
[0091] Moreover, in contrast to both of the embodiments outlined above, in those exemplary embodiments that only extract meaning from the combination of the recognized multimodal inputs, the multimodal parser/meaning recognition system 300 does not generate the new grammar or language model lattice 160. Thus, the gesture recognition lattice 255 does not provide compensation to the automatic speech recognition system 100. Rather, the multimodal parser/meaning recognition system 300 only combines the gesture recognition lattice 255 and the recognized possible word sequences lattice 175 to generate the recognition meaning lattice 375.
[0092] When the gesture recognition system 200 generates only a single recognized possible sequence of gestures as the gesture recognition lattice 255, that means there is essentially no uncertainty in the gesture recognition. In this case, the gesture recognition lattice 255 provides compensation to the automatic speech recognition system 100 for any uncertainty in the speech recognition process. However, the gesture recognition system 200 can generate a lattice of n possible recognized gesture sequences as the gesture recognition lattice 255. This recognizes that there may also be uncertainty in the gesture recognition process.
[0093] In this case, the gesture recognition lattice 255 and the word lattice 155 provide mutual compensation for the uncertainties in both the speech recognition process and the gesture recognition process. That is, in the face of this uncertainty, the best, i.e., most-probable, combination of one of the n-best word sequences in the word lattice 155 with one of the n-best gesture sequences in the gesture recognition lattice may not include the best recognition possible sequence from either the word lattice 155 or the gesture recognition lattice 255. For example, the most-probable sequence of gestures in the gesture recognition lattice may combine only with a rather low-probability word sequence through the word lattice, while the most-probable word sequence may combine well only with a rather low-probability gesture sequence. In contrast, a medium-probability word sequence may match very well with a medium-probability gesture sequence. Thus, the net probability of this latter combination of word and gesture sequences may be higher than the probability of the combination of the best word sequence with any of the gesture sequences through the gesture recognition lattice 255 and may be higher than the probability of the combination of the best gesture sequence with any of the word sequences through the lattice of possible word sequences 155. In this way, mutual compensation is provided between the gesture recognition system 200 and the automatic speech recognition system 100.
[0094] FIGS. 3-5 illustrate in greater detail various exemplary embodiments of the gesture recognition system 200, the multimodal parser/meaning recognition system 300, and the multimodal user input device 400. In particular, as shown in FIG. 3, one exemplary embodiment of the gesture recognition system 200 includes a gesture feature extraction subsystem 210 and a gesture recognition subsystem 230. Various other exemplary embodiments may include a gesture language model lattice and a gesture meaning subsystem. In operation, gesture utterances are input through the gesture input portion 410 of the multimodal user input device 400, which converts the movements of an input device, such as a mouse, a pen, a trackball, a track pad or any other known or later-developed gestural input device, into an electronic gesture signal 205. At the same time, the multimodal user input device 400 converts the gestural input into digital ink that can be viewed and understood by the user. This is shown in greater detail in FIG. 5.
[0095] The gesture feature extraction subsystem 210 converts the motions of the gesture input device represented by the gesture signal 205 into a gesture feature lattice 220. As disclosed in Johnston 1-3, the various gestures that can be made can be as simple as pointing gestures to a particular information element at a particular location within the gesture input portion 410 of the multimodal user input device 400, or can be as complex as a specialized symbol that represents a type of military unit on a military map displayed in the gesture input portion 410 of the multimodal user input portion 400 and includes an indication of how the unit is to move, and which unit is to move and how far that unit is to move, as described in detail in Johnston 1.
[0096] The gesture feature lattice 220 is input to the gesture recognition subsystem 230. The gesture recognition subsystem 230 may be implemented as a neural network, as a Hidden-Markov Model (HMM) or as a simpler template-based gesture classification algorithm. The gesture recognition subsystem 230 converts the gesture feature lattice 220 into the gesture recognition lattice 255. The gesture recognition lattice 255 includes the identities of graphical elements against which diectic and other simple “identification” gestures are made, possible recognition of more complex gestures that the user may have made and possibly the locations on the displayed graphics where the more complex gesture was made, such as in Johnston 1, and the like. As shown in FIG. 2, the gesture recognition system 200 outputs the gesture recognition lattice 255 to the multimodal parser/meaning recognition system 300.
[0097] It should be appreciated that the gesture feature recognition subsystem 210 and the gesture recognition subsystem 230 can each be implemented using any known or later-developed system, circuit or technique that is appropriate. In general, the entire gesture recognition system 200 can be implemented using any known or later-developed system that generates a directed graph from a gesture input.
[0098] For example, one known system captures the time and location or locations of the gesture. Optionally, these inputs are then normalized and/or rotated. The gestures are then provided to a pattern classification device that is implemented as part of the gesture feature recognition subsystem 210. In various exemplary embodiments, this pattern classification device is a template matching system, which transforms the gesture into a feature vector. In various other exemplary embodiments, this pattern classification device is a neural network or a Hidden Markov Model that has been trained to recognize certain patterns of one or more temporally and/or spatially related gesture components as a specific set of features.
[0099] When a single gesture is formed by two or more temporally and/or spatially related gesture components, those gesture components can be combined into a single gesture either during the recognition process or by the multimodal parser/meaning recognition system 300. Once the gesture features are extracted, the gesture recognition subsystem 230 combines the temporally adjacent gestures into a lattice of one or more recognized possible gesture sequences that represent how the recognized gestures follow each other in time.
[0100] In various exemplary embodiments, the multimodal parser and meaning recognition system 300 can be implemented using a single three-tape finite-state device that inputs the output lattices from the speech recognition system 100 and the gesture recognition system 200 and directly obtains and outputs a meaning result. In various exemplary embodiments, the three-tape finite-state device is a three-tape grammar model that relates the gestures and the words to a meaning of the combination of a gesture and a word. FIG. 7 shows a portion of such a three-tape grammar model usable in the multimodal parser and meaning recognition system 300 to generate a meaning output from gesture and speech recognition inputs. In general, the multimodal parser and meaning recognition system 300 can be implemented using an n-tape finite-state device that inputs n−1 lattices from a plurality of recognition systems usable to recognize an utterance having a plurality of different modes.
[0101] FIG. 4 shows the multimodal parser/meaning recognition system 300 in greater detail. As shown in FIG. 4, the multimodal parser/meaning recognition system 300 may include one or more of a gesture-to-speech composing subsystem 310, a gesture-to-speech finite-state transducer 320, a lattice projection subsystem 330, a gesture and speech composing subsystem 340, a speech/gesture combining subsystem 350, a speech/gesture/meaning lattice 360 and/or a meaning recognition subsystem 370. In particular, the gesture-to-speech composing subsystem 310 inputs the gesture recognition lattice 255 output by the gesture recognition system 200 and composes it with the gesture-to-speech finite-state transducer 320 to generate a gesture/language finite-state transducer 325. The gesture/language finite-state transducer 325 is output to both the lattice projection subsystem 330 and the gesture and speech composing subsystem 340.
[0102] The lattice projection subsystem 330 generates a projection of the gesture/language finite-state transducer 325 and outputs the projection of the gesture/language finite-state transducer 325 as the grammar or language model lattice 160 to the automatic speech recognition system 100. Thus, if the multimodal parser/meaning recognition system 300 does not also extract meaning, the gesture and speech composing subsystem 340, the speech/gesture combining subsystem 350, the speech/gesture/meaning lattice 360 and the meaning recognition subsystem 370 can be omitted. Similarly, if the multimodal parser/meaning recognition system 300 does not generate a new grammar or language model lattice 160 for the automatic speech recognition system 100, at least the lattice projection subsystem 330 can be omitted.
[0103] In those various embodiments that combine the gesture recognition lattice 255 and the recognized possible lattice of word sequences 175, whether or not the automatic speech recognition 100 has generated the lattice of possible word sequences 175 based on using the projection of the gesture/language finite-state transducer 325 as the grammar or language model or lattice 160, the lattice of possible word sequences 175 is input by the multimodal parser/meaning recognition system 300. In particular, the gesture and speech composing subsystem 340 inputs both the lattice of possible word sequences 175 and the gesture/language finite-state transducer 325. In those various exemplary embodiments that do not use the output of the gesture recognition system 200 to provide compensation between the speech and gesture recognition systems 100 and 200, the gesture/language finite-state transducer 325 can be generated using any known or later-developed technique for relating the gesture recognition lattice 255 to the recognized possible lattice of word sequences 175 in place of the gesture-to-speech composing subsystem 310 and the gesture-to-speech finite-state transducer 320.
[0104] In those various exemplary embodiments that extract meaning from the multimodal inputs, the gesture and speech composing subsystem 340 composes these lattices to generate a gesture/speech finite-state transducer 345. The gesture and speech composing subsystem 340 outputs the gesture/speech finite-state transducer 345 to the speech/gesture combining subsystem 350. The speech/gesture combining subsystem 350 converts the gesture/speech finite-state transducer 345 to a gesture/speech finite-state machine 355. The gesture/speech finite-state machine 355 is output by the speech/gesture combining subsystem 350 to the meaning recognition subsystem 370. The meaning recognition subsystem 370 composes the gesture/speech finite-state machine 355 with the speech/gesture/meaning finite-state transducer 360 to generate a meaning lattice 375. The meaning lattice 375 combines the recognition of the speech utterance input through the speech input portion 420 and the recognition of the gestures input through the gesture input portion 410 of the multimodal user input device 400. The most probable meaning is then selected from the meaning lattice 375 and output to a downstream task.
[0105] To illustrate one exemplary embodiment of the operation of the multimodal recognition and/or meaning system 1000, the multimodal user input device 400 includes, for example, the gesture input portions 410 and the speech input portion 420 shown in FIG. 5. In particular, the gesture input portion 410 displays a graphical user interface that allows the user to direct either e-mail messages or pager messages to the various persons, departments, and/or organizations represented by the objects 412 displayed in the gesture input portion 410. The multimodal user input device 400 also allows the user to input spoken commands to the speech input portion, or microphone, 420. For simple illustration, further assume that the user has generated the two gestures 414 shown in FIG. 5 and has spoken the utterance “e-mail this person and that organization” in association with generating the gestures 414 against the graphical user interface object 412 labeled “Robert DeNiro” and the graphical user interface object 412 labeled “Monumental Pictures”, respectively.
[0106] The structure interpretation of multimodal commands of this kind can be captured declaratively in a multimodal context-free grammar. A multimodal context-free grammar can be defined formally as the quadruple MCFG as follows: 1 TABLE 1 MCFG = <N, T, P, S > where N is the set of non-terminals; P is the set of projections of the form: A→&agr; where A∈N and &agr;∈(N∪T)*; S is the start symbol for the grammar; T is the set of terminals: ((W∪&egr;) × (G∪&egr;) × (M∪&egr;)+), where W is the vocabulary of the speech; G is the vocabulary of gesture: G = (GestureSymbols ∪ EventSymbols); GestureSymbols = {Gp, Go, Gpf, Gpm . . .}; Finite collections of EventSymbols = {e1, e2 . . . }; and M is the vocabulary that represents meaning and includes EventSymbols⊂M.
[0107] In general, a context-free grammar can be approximated by a finite-state automaton. The transition symbols of the finite-state automaton are the terminals of the context-free grammar. In the case of the multimodal context-free grammar defined above, these terminals contain three components, W, G and M. With respect to the discussion outlined above regarding temporal constraints, more specific temporal constraints than order can be encoded in the finite-state approach by writing symbols representing the passage of time onto the gesture tape and referring to such symbols in the multimodal grammar.
[0108] In the exemplary embodiment of the gesture input portion 410 shown in FIG. 5, the gestures 414 are simple deictic circling gestures. The gesture meaning subsystem 250 assigns semantic types to each gesture 414 based on the underlying portion of the gesture input portion 410 against which the gestures 414 are made. In the exemplary embodiment shown in FIG. 5, the gestures 414 are made relative to the objects 412 that can represent people, organizations or departments to which an e-mail message or a pager message can be directed. If the gesture input portion 410 were instead a map, the gestures would be referenced against particular map coordinates, where the gesture indicates the action to be taken at particular map coordinates or the location of people or things at the indicated map location.
[0109] In Johnston 4 and the incorporated 253 application compared with a feature-based multimodal grammar, these semantic types constitute a set of atomic categories which make the relevant distinctions for gesture events to predict speech events and vice versa. For example, if the gesture is a deictic, i.e., pointing, gesture to an object in the gesture input portion 410 that represents a particular person, then spoken utterances like “this person”, “him”, “her”, and the like, are the preferred or predicted speech events and vice versa. These categories also play a role in constraining the semantic representation when the speech is underspecified with respect to the semantic type, such as, for example, spoken utterances like “this one”.
[0110] In Johnston 4 and the incorporated 253 application, the gesture symbols G can be organized into a type hierarchy reflecting the ontology of the entities in the application domain. For example, in the exemplary embodiment of the gesture input portion 410 shown in FIG. 5, a pointing gesture may be assigned the general semantic type “G”. This general semantic gesture “G” may have various subtypes, such as “Go” and “Gp”, where “Go” represents a gesture made against an organization object, while the “Gp” gesture is made against a person object. Furthermore, the “Gp” type gesture may itself have subtypes, such as, for example, “Gpm” and “Gpf” for objects that respectively represent male and female persons.
[0111] The systems and methods for recognizing and representing gestures disclosed in U.S. patent application (Attorney Docket No. 110805), incorporated herein by reference in its entirety, provide an approach that can be used instead of, or in addition, to that discussed above and shown in FIGS. 1-5. Instead of, or in addition to, using atomic symbols in the multimodal grammar and the corresponding finite-state machines to represent different types of gestures, gesture interpretations are encoded as sequences of symbols. In this approach, each symbol conveys a specific attribute of the content of the gesture, such as type or number. Under this approach, the Gesture Symbols and Event Symbols in the multimodal grammar fragment, shown in Table 1 are instead defined as: 2 TABLE 2 GestureSymbols = {G, area, location, restaurant, 1, . . . }; EventSymbols = {SEM}
[0112] These definitions can be used instead of, or in addition to, the definitions of the Gesture Symbols and Event Symbols shown in Table 1.
[0113] FIG. 6 illustrates a fragment of a context-free grammar that is capable of handling a set of gestures and a set of spoken utterances that would be recognizable using the finite state machines shown in FIGS. 12 and 13. FIG. 7 illustrates a three-tape finite state automaton corresponding to the multimodal context-free grammar fragment shown in FIG. 6.
[0114] By decomposing the gesture symbols into sequences of symbols, it is easier to reference sets of entities of a specific type. In addition, a smaller number of symbols are required in the alphabet of symbols that represent the gestural content of the grammar. Further, decomposing the gesture symbols into sequences of symbols facilitates storing specific gesture content, discussed below, and aggregating adjacent selection gestures, as disclosed in U.S. patent application (Attorney Docket No. 112364).
[0115] Under this approach, for example, the gesture symbol complexes have a basic format such as:
[0116] G FORM MEANING (NUMBER TYPE) SEM
[0117] However, it should be appreciated that the gesture symbol complexes can be implemented in any appropriate format. The “FORM” term specifies the physical form of the gesture. In various exemplary embodiments, the “FORM” term can take such values as area, point, line and arrow. The “MEANING” term specifies the specific meaning of the form of the gesture. For example, if the value for the “FORM” term of a gesture is area, the value for the “MEANING” term of that “FORM” term can be location, selection or any other value that is appropriate. If the value of the “MEANING” term of that “FORM” term is selection, such that one or more specific entities are selected, the “NUMBER” term and the “TYPE” term can be used to further specify the entities that are or have been selected. In particular, the value for the “NUMBER” term specifies the number of entities selected, such as 1, 2, 3, “many” and the like. Similarly, the value for the “TYPE” term specifies a particular type of entity, such as restaurant, theatre and the like, as appropriate for the given implementation or use. A value of mixed can be used for the “TYPE” term when one or more associated gestures reference a collection of entities of different types. The “SEM” term is a place holder for the specific content of the gesture, such as the points that make up an area or the identifiers (ids) of objects in a selection. To facilitate recomposing specific gestural content, specific content is mapped to a distinguished symbol, such as, the “SEM” term, while the other attributes of the gesture are mapped to themselves.
[0118] When using finite-state methods, in order to capture multimodal integration, it is necessary to abstract over specific aspects of gestural content. In the systems and methods disclosed in U.S. patent application (Attorney Docket No. 112365, which is filed on even date herewith), incorporated herein by reference in its entirety, abstraction is performed by representing the input gesture as a finite-state transducer that maps the specific contents of the input gesture to a distinguished symbol, such as, for example, the “SEM” term, while the other attributes of the input gesture are mapped to themselves. To perform multimodal integration, the output projection of the gesture finite-state transducer is used. After multimodal integration is completed, a projection of the gesture input and the meaning portions of the resulting finite-state machine is taken. The projection of the resulting finite-state machine is composed with the gesture input finite-state transducer to reintegrate the specific content of the gesture input which was left out of the finite-state process.
[0119] Thus, in the finite-state approach used in the systems and methods disclosed in incorporated (Attorney Docket No. 112365) application, the specific content of the input gesture is essentially stored in the gesture finite-state transducer and a projection of the output of the gestural input finite-state transducer is used to conduct multimodal modal integration using finite-state devices. Multimodal integration is performed and a resulting finite-state device, which relates the gesture input and at least one other mode of multimodal input to meaning, is generated. After multimodal integration is performed, a projection of the resulting finite-state device is taken such that the projection contains the gesture input and the meaning portions of the resulting finite-state device. The specific content of the input gesture is then retrieved from the gesture finite-state transducer by composing the gesture finite-state transducer with the projection of the resulting finite-state device.
[0120] FIGS. 12 and 13 are examples of transducers I:G, where G contains gesture symbols and I contains the gesture symbols and specific contents of the gesture. In particular, the G side contains information identifying the type of gesture and its properties and the I side contains the information on the G side, however, where the G side contains a place holder for the specific contents of the gesture, the I side contains the specific contents. Therefore, I and G differ only in cases where the gesture symbol on G is SEM. If the symbol on G is SEM, then I contains the specific content of the gesture. Accordingly, when there is specific content, such as a list of entities or points in the gesture stream, the symbol in G is the reserved symbol SEM and the specific content is placed on the I side opposite the SEM.
[0121] In addition, in one exemplary embodiment of the systems and methods disclosed in the incorporated (Attorney Docket No. 112365) application, a projection, i.e., the gesture input finite-state machine, is then used to perform multimodal integration. Accordingly, in order to reintegrate the specific content of the input gesture after multimodal integration is performed, an output projection of the resulting meaning finite-state machine is composed with the gesture finite-state transducer.
[0122] Further, in a multimodal system that integrates gestures and speech, deictic numeral expressions, such as, for example, “these four restaurants” can refer to a multitude of different possible sequences of gestures that are compatible and should be integrated with the spoken phrase. For example, the user might circle four individual restaurants, a set of four restaurants, two sets of two restaurants or a set of three restaurants and a single restaurant, etc. Therefore, in a multimodal recognition system that integrates gestures and speech, for example, capturing all of the gestural possibilities in the spoken language grammar significantly increases the size and complexity of the finite state devices used to represent the multimodal recognition. Thus, for such a known multimodal recognition system, every time a numerical expression exists in a multimodal input, a multitude of the alternative multimodal subcategorization frames would need to be assigned to that numerical expression by that known multimodal recognition system. Thus, any plural expression can be massively ambiguous.
[0123] This problem is presented in “Deixis and Conjunction in Multimodal Systems”, M. Johnston, Proceedings of COLING 2000, Saarbruecken, Germany, 2000 (Johnston 5), herein incorporated by reference in its entirety. In Johnston 5, the problem is addressed by underspecifying the configuration of gestures in the multimodal subcategorization of the deictic numeral expressions. For example, instead of subcategorizing for a sequence of N gestures, the deictic numeral expression, ‘these N’, for example, subcategorizes for a collection of plurality N:<G[number:N]>. For example, the expression ‘these four people’ is subcategorized into <Gperson[number:4]>. Gesture combination rules enumerate the different ways to refer to a collection of entities.
[0124] In the systems and methods disclosed in Johnston 4 and the incorporated 253 application, certain simplifying assumptions are used with respect to temporal constraints in a multimodal recognition system that integrates gestures and speech. For multi-gesture utterances, the primary function of temporal constraints is to force an order on the gestures. For example, if a user generates the spoken utterance “move this here” and simultaneously makes two gestures, then the first gesture corresponds to the spoken utterance “this”, while the second gesture corresponds to the spoken utterance “here”.
[0125] In the various exemplary embodiments of the systems and methods according to the invention described in Johnston 4 and the incorporated 253 application, the multimodal grammars encode order, but do not impose explicit temporal constraints. In Johnston 4 and the incorporated 253 application, the deictic numerals are subcategorized for a plurality of the appropriate number and a set of gesture combination rules are predictively applied to combine the elements of the gestural input into the appropriate pluralities.
[0126] In the systems and methods according to this invention, aggregation is used to pre-process the input lattice. For example, “gesture aggregation” is used to pre-processes the gesture lattice. Aggregation adds one or more additional branches that represent combinations of related inputs to the lattice. For example, the inputs may be related spatially and/or temporally and can represent entities of the same or different type. The aggregated branch can then be combined with another input mode of the multimodal device, such as, for example, speech input that refers to the combination of the input gestures represented by the aggregated branch.
[0127] Aggregating temporally-related and/or spatially-related inputs according to the systems and methods of this invention significantly simplifies the construction of the multimodal grammar. The multimodal grammar can be significantly simplified by reducing, and ideally eliminating, the need to include all the possible combinations that could be assigned to such inputs, such as, for example, deictic numeral expressions. Accordingly, in various exemplary embodiments, the systems and methods of this invention also allow the size the finite state devices which are used to represent the multimodal grammar to be reduced, sometimes significantly. For example, in various exemplary embodiments, the finite state devices which are used for multimodal integration will often be significantly smaller than they otherwise would have been using this known technique:
[0128] For example, using the gesture symbol representation scheme disclosed in U.S. patent application (Attorney Docket No. 110805, filed on even date herewith), incorporated herein by reference in its entirety, an expression such as “these three restaurants” is assigned, for example, the gesture stream “G area selection 3 restaurant SEM” by the multimodal grammar. The spoken utterance stream will combine with a single area gesture containing three restaurants. However, to allow this single spoken utterance to readily combine with a sequence of three temporally-related but spatially-separated gestures, for example, aggregation should first be applied. In this example, aggregation is used to preprocess the gesture lattice by parsing the gesture lattice and composing together temporally-adjacent selections of identical type. Additional branches, representing the composed selections, are added to the gesture lattice.
[0129] For example, if the user makes three gestures, such as, “G area selection 1 restaurant SEM, G area selection 1 restaurant SEM and G area selection 1 restaurant SEM”, the resulting gesture lattice could include an aggregation of all three gestures and/or an aggregation of each temporally-related and/or spatially-related pair.
[0130] In various exemplary embodiments of the systems and methods according to this invention, it is also possible to maintain a record of the original inputs, i.e., a record of the non-aggregated inputs. A record of the original sequence of inputs can be preserved, for example, by assigning weights to the paths in the resulting aggregated lattice. For example, in various exemplary embodiments, the newly aggregated paths can be assigned lower weights than the paths representing the original input sequences. Alternatively, for example, in various exemplary embodiments, a record of the original sequence of inputs can be maintained by adding a symbol, such as, for example, the symbol ‘AGG’, to the multimodal grammar and the resulting aggregated paths in the finite-state device. In addition, in various exemplary embodiments of the systems and methods according to this invention, the grammar may also be set up to aggregate only certain expressions.
[0131] In the finite-state automata approach used in the systems and methods according to this invention, in addition to capturing the structure of language with the finite-state device, meaning is also captured. This is significant in multimodal language processing, because the central goal is to capture how the multiple modes contribute to the combined interpretation. In the finite-state automata technique used in the systems and methods according to this invention, symbols are written onto the third tape of the three-tape finite-state automaton, which, when concatenated together, yield the semantic representation for the multimodal utterance.
[0132] FIG. 9 illustrates a second exemplary embodiment and application of a multimodal recognition and/or meaning system 1000 which uses the systems and methods for recognizing and representing gesture according to this invention. For example, in this second exemplary embodiment, the user interacts with a graphical interface displaying restaurant listings and a dynamic map showing locations and street information. The gesture input portion 410 can be used to display, for example, a graphical user interface that includes a working-city-guide application and navigation screen that, for example, enables mobile users to access restaurant and subway information for a city. In this example, the user is able to interact with the graphical interface to display restaurant listings and a dynamic map showing locations and street information. This graphical user interface responds to commands entered using pen-based gestures and/or spoken utterances.
[0133] The user is, for example, free to give commands or to reply to requests displayed on the graphical user interface using speech, by drawing on the display with a stylus, or using synchronous multimodal combinations of the available modes. The user can, for example, ask for the review, cuisine, phone number, address, or other information for a restaurant or set of restaurants. The working-city-guide application generates graphical callouts on the display.
[0134] For example, a user can request to see restaurants using the spoken command “Show cheap Italian restaurants in Chelsea”. The working-city-guide application will then display the appropriate map location and show the locations of the restaurants that meet the indicated criteria. Alternatively, the user can give the same command multimodally by circling an area on the map and uttering “Show cheap Italian restaurants in this neighborhood”. As shown in FIG. 10, if the immediate environment is too noisy or public, the same command can be input solely using a writing gesture, by circling an area of the displayed map and by writing “cheap” and “Italian”.
[0135] As shown in FIG. 11, for example, the user can speak the utterance “Phone numbers for these restaurants” and circle, for example, a total of three restaurants. As shown in FIG. 12, a callout with the restaurant name and number, such as for example, the text string, “Le Zie can be reached at 212-206-8686”, is then displayed for one or more of the identified restaurants. Alternatively, any of the information seeking commands can also be input solely using gestures. For example, the user could alternatively have circled the restaurants and written phone. The user can also pan and zoom around the map using any of a variety of input modes. For example, the user can speak the utterance “show upper west side” or circle an area and speak the utterance “zoom in here”.
[0136] FIG. 14 shows one exemplary embodiment of a resulting gesture lattice that is obtained when the user draws an area on the screen which contains two restaurants, that have identifiers id1 and id2. The arc labels of the gesture lattice are the symbols which define the gesture symbol complex. Further, if the user utters the spoken string “Show me Chinese restaurants in this neighborhood”, then the top path, comprising states 0-2 and states 3-7, of the gesture lattice shown in FIG. 14 will be chosen when the multimodal finite-state device is applied. In contrast, if the user utters the spoken string “Tell me about these two restaurants”, then the lower path, comprising states 0-2 and states 4-7, of the gesture lattice shown in FIG. 14 will be chosen.
[0137] FIG. 13 shows one exemplary embodiment of a resulting gesture lattice that is obtained when the user circles a restaurant and a theatre. In contrast, if the user utters the spoken string, “Tell me about this theatre”, the middle path, comprising states 0-2, state 4, state 5, state 7 and state 10, of FIG. 13 will be chosen. Similarly, if the user utters the spoken string, “Tell me about these two”, the bottom path, comprising the states 0-2, state 4 and states 8-10, of FIG. 13 will be chosen. In particular, the mixed path, is used when the user circles several entities and selects a specific one of the circled entities by type.
[0138] Accordingly, by splitting the symbols into symbol complexes, it is not necessary to have a unique symbol for “G_area_location”, “G_area_selection—1_restaurant”, “G_area_selection—2_restaurant”, and the like, as would be required when using only the just first approach shown in FIGS. 1-5. Therefore, the number of required symbols in the alphabet of the gesture symbols can be reduced and possibly significantly reduced. In addition, in various exemplary embodiments, general categories are separately represented in the symbol complexes. Thus, in various exemplary embodiments, it is easier to reference to sets of entities of a specific. Accordingly, in various exemplary embodiments of the systems and methods for recognizing and representing gestures according to this invention, if the word place, in the spoken utterance, “Tell me about this place”, referred to either a restaurant or a theatre, the word “place” could be associated with both arcs in the gesture lattice. In addition, splitting gesture symbols also facilitates in later aggregating the symbols, as well as in recovering specific gestural content in systems and methods which use finite state machines.
[0139] In various exemplary embodiments, the working-city-guide application can also provide a summary, one or more comparisons, and/or recommendations for an arbitrary set of restaurants. The output is tailored to the user's preferences based on a user model, which is, for example, based on answers to a brief questionnaire. For example, as shown in FIG. 15, the user could speak the utterance “compare these restaurants” and circle a large set of restaurants. In various exemplary embodiments, a number, such as, for example, the top three restaurants are then ranked according to the user model. In various exemplary embodiments, if the user is more concerned about price than food quality, the working-city-guide application would, for example, order the selected restaurants by weighting price more highly than food quality in analyzing the selected restaurants.
[0140] The working-city-guide application can also provide subway directions. For example, if the user speaks the spoken utterance “How do I get to this place?” and circles one of the restaurants displayed on the map, the working-city-guide application will, for example, display a query or graphical callout, such as, for example, the text string “Where do you want to go from?” The user can then respond, for example, with a spoken utterance stating a location, by saying a location, such as, for example, “25th Street and 3rd Avenue”, with a gesture defining a location, for example, by writing “25th Street and 3rd Avenue”, or by pointing to 25th Street and 3rd Avenue on the map, or multimodally by speaking the spoken utterance “from here” and pointing to or circling the location on the map.
[0141] The working-city-guide application then determines a subway route. In various exemplary embodiments, as appropriate, the working city-guide-application generates a multimodal presentation indicating the series of actions the user needs to take will be displayed. In various exemplary embodiments, the working-city-guide application starts, for example, by zooming in on the first station and then gradually presenting each stage of the route along with a series of synchronized test-to-speech (TTS) prompts. FIG. 16 shows one exemplary embodiment of a resulting subway route.
[0142] It should be thus be appreciated, from the preceding description of Table 2 and FIGS. 6-16, that the interpretations of electronic ink may be encoded as symbol complexes. In various exemplary embodiments the symbol complexes can have, for example, the form “G FORM MEANING (NUMBER TYPE) SEM”. FIG. 17 shows, for example, the resulting gesture lattice, including the additional aggregated paths as discussed below, of the gesture made in FIG. 11 using this exemplary embodiment of the symbol complexes according to this invention.
[0143] As discussed above, the user can speak the utterance “Phone numbers for these three restaurants” and circle, for example, as shown in FIG. 11, a total of three restaurants. As shown in FIG. 11, the user may combine the diectic spoken expression, “these three restaurants”, for example, with two area gestures, one which selects one restaurant and one which selects the other two restaurants. Alternatively, the user may, for example, combine the deictic spoken expression, “these three restaurants” with three separate selection gestures. FIG. 17 shows one exemplary embodiment of a gesture lattice that can result from these possibilities. The first gesture (states 0-7) is either a reference to a location (states 0-3 and 7) or a reference to a restaurant (states 0-2 and 4-7). The second gesture (states 7-13 and 16) is either a reference to a location (states 7-10 and 16), or to a set of two restaurants (states 7-9, 11-13 and 16).
[0144] In a multimodal recognition system that integrates gestures and speech, to combine a deictic spoken utterance with multiple gestural inputs according to the systems and methods of this invention, the gesture lattice is pre-processed and spatially-related and/or temporally-related gestures are composed. The related gestures may, for example, be selections of identical type, such as, for example, a plurality of restaurant selections. Alternatively, depending on the spoken utterance, for example, the related gestures may, for example, be selections of mixed type, such as, for example, a restaurant and a theatre selection. Based on the compositions, branches are added to the gesture lattice. FIG. 17 shows, for example, the resulting gesture lattice including the additional paths resulting from aggregation. The aggregation process applies to the two adjacent selections and adds a selection of three restaurants (states 0-2, 4 and 14-16).
[0145] However, in a multimodal recognition system that integrates gestures and speech, if, for example, the user says “show me Chinese restaurants in this neighborhood and this neighborhood”, the path containing the two locations (states 0-3, 7-10 and 16) will be taken when the gesture lattice is combined with the speech lattice. Similarly, if, for example, the user says “tell me about this place and these places”, then the path containing states 0-2, 4-9, 11-13 and 16, with the adjacent selections, is taken when the gesture lattice is combined with the speech lattice. In addition, if the user says “tell me about these or phone numbers for these three restaurants”, then the aggregate path containing states 0-2, 4 and 14-16 will be taken.
[0146] It should be appreciated that the systems and methods according to this invention apply aggregation to combine gestures or inputs that are of identical type, i.e., type-specific aggregation. That is, type-specific aggregation refers to combining multiple gestures, such as for example, multiple gestures referring only to restaurants or multiple gestures referring only to theatres.
[0147] However, the systems and methods of this invention may also be used to combine gestures or inputs of mixed type, i.e., non-type-specific aggregation. That is, non-type-specific aggregation refers to combining multiple gestures, such as, for example, a gesture referring to one type of entity with another gesture referring to a second type of entity. For example, one gesture may refer to a restaurant, while a second gesture may refer to a theatre. For example, if the user says “tell me about these two” and circles a restaurant and a theatre, non-type-specific aggregation applies to combine the two gestures into an aggregate gesture or path of mixed type. For example, an aggregated path that can combine with the speech “these two”, can be: “G:G area:area selection:selection 2:2 mixed:mixed ([id4, id5]):SEM”.
[0148] In addition, the systems and methods according to this invention also include systems and methods for preserving an indication of the original sequences of the gestures. For example, in various exemplary embodiments, it is possible to preserve the original sequence of the gestures by assigning lower weights to the paths that were added as a result of aggregation.
[0149] Further, in order to abstract over specific aspects of the gestural content in accordance with the systems and methods of this invention, the gesture lattice is converted to a gesture transducer I:G, where the G side is the set of gesture symbols (including SEM) and I contains both gesture symbols and the specific contents. The specific contents of the gestural input includes, for example, entities or points on the gesture input portion 410 that are selected by the user. Further, the specific content may, for example, be an identification number, i.e., an entity identifier, that represents the entity or point on the gesture input portion. For example, to convert the gesture lattice shown in FIG. 17 to a gesture transducer I:G, for example, the path, comprising states 15-16 and reciting SEM([id1, id2, id3]) in FIG. 17 becomes [id1, id2, id3]:SEM.
[0150] Additionally, a gesture transducer I:G may be generated by the gesture recognition system based on the gestural input. In this case, the gesture transducer is converted to a gesture finite-state machine usable to carry out multimodal integration or any other applicable function using finite-state devices. For example, if the user circles two restaurants and the user says “phone numbers for these two restaurants” the gesture is represented as a transducer as shown in FIG. 13, where the transducer I:G has same symbol on each side except for the SEM arcs, which are split. In this case, to carry out the multimodal integration, or any other function performed with finite-state devices, a projection of the output of the transducer, that is a projection of the gesture path G is taken. Accordingly, the projection of the output of the gesture transducer is used to perform the applicable function using finite-state devices.
[0151] After the gesture symbols G and the words W are integrated using the finite-state devices G:W and G_W:M, for example, i.e., after multimodal integration, the gesture path G and meaning path M in the resulting finite-state device are used to re-establish the connection between the SEM symbols and their specific contents, for example entities or points selected by the user, that are stored in the I path of the gesture transducer I:G. In particular, in order to reintegrate the specific contents of the gesture, the gesture transducer I:G is composed with the gesture path G and meaning path M of the device resulting from multimodal integration (I:G o G:M=I:M). In addition, in order to output the meaning, the symbols on the M side are concatenated together. Further, when outputting the meaning, if the M symbol is SEM, the symbol on the I side is taken for that arc.
[0152] While a three-tape finite-state automaton is feasible in principle, currently available tools for finite-state language processing generally only support two-tape finite-state automata, i.e., finite-state transducers. Furthermore, speech recognizers typically do not support the use of a three-tape finite-state automaton as a language model. Accordingly, the multimodal recognition and/or meaning system 1000 implements this three-tape finite-state automaton approach by using a series of finite-state transducers in place of the single three-tape finite-state automaton shown in FIG. 18, as described below. In particular, the three-tape finite-state automaton shown in FIG. 18 and illustrated by the grammar fragment shown in FIG. 19 can be decomposed into an input component relating the gesture symbols G and the word symbols Wand an output component that relates the input component to the meaning symbols M.
[0153] FIG. 18 shows a three-tape finite-state automaton that corresponds to the grammar fragment shown in FIG. 19 and that is usable to recognize the meaning of the various spoken and gestural inputs that can be generated using the exemplary graphical user interfaces displayed using the gesture input portion 410 of the multimodal user input devices 400 shown in FIGS. 5 and 9. The three-tape finite-state automaton shown in FIG. 18 is decomposed into the gesture-to-speech finite-state transducer shown in FIG. 20 and the speech/gesture/meaning finite-state transducer shown in FIG. 21.
[0154] The gesture-to-speech finite-state transducer shown in FIG. 20 maps the gesture symbols G to the word symbols W that are expected to coincide with each other. Thus, in the exemplary embodiment of the multimodal user input device 400 shown in FIG. 5, the verbal pointers “that” and “this” are expected to be accompanied by the deictic gestures 414 made against either a department object, an organization object or a person object 412.
[0155] The gesture-to-speech transducer shown in FIG. 20 captures the constraints that the gestures, made by the user through the gesture input portion 410 of the multimodal user input device 400, place on the speech utterance that accompanies those gestures. Thus, a projection of the output tape of the gesture-to-speech finite-state transducer shown in FIG. 20 can be used, in conjunction with the recognized gesture string, such as the recognized gesture string shown in FIG. 23 that represents the gestures illustrated in the exemplary embodiment of the multimodal user input device 400 shown in FIG. 9, as a language model usable to constrain the possible sequences of words to be recognized by the utterance recognition subsystem 170 of the automatic speech recognition system 100.
[0156] It should be appreciated that, in those exemplary embodiments that do not also extract meaning, the further processing outlined below with respect to FIGS. 20-32 can be omitted. Similarly, in those exemplary embodiments that do not use one or more of the multimodal inputs to provide compensation to one or more of the other multimodal inputs, the processing outlined above with respect to FIGS. 20, 21 and 26 can be omitted.
[0157] The speech/gesture/meaning finite-state transducer shown in FIG. 21 uses the cross-product of the gesture symbols G and the word symbols W as an input component or first tape. Thus, the gesture-to-speech finite-state transducer shown in FIG. 20 implements the function : G→W. The output or second tape of the speech/gesture/meaning finite-state transducer shown in FIG. 21 contains the meaning symbols M that capture the semantic representation of the multimodal utterance, as shown in FIG. 18 and outlined above. Thus, the speech/gesture/meaning finite-state transducer shown in FIG. 21 implements the function ℑ: (G×W)→M. That is, the speech/gesture/meaning finite-state transducer shown in FIG. 21 is a finite-state transducer in which gesture symbols and words are on the input tape and the meaning is on the output tape.
[0158] Thus, the gesture-to-speech finite-state transducer and the speech/gesture/meaning finite-state transducers shown in FIGS. 20 and 21 are used with the speech recognition system 100 and the multimodal parser/meaning recognition system 300 to recognize, parse, and/or extract the meaning from the multimodal inputs received from the gesture and speech input portions 410 and 420 of the multimodal user input device 400.
[0159] It should be appreciated that there are any variety of ways in which the multimodal finite-state transducers can be integrated with the automatic speech recognition system 100, the gesture recognition system 200 and the multimodal parser/meaning recognition system 300. Clearly, for any particular recognition task, the more appropriate approach will depend on the properties of the particular multimodal user input interface 400 through which the multimodal inputs are generated and/or received.
[0160] The approach outlined in the following description of FIGS. 18-32 involves recognizing the gesture string first. The recognized gesture string is then used to modify the language model used by the automatic speech recognition system 100. In general, this will be appropriate when there is limited ambiguity in the recognized gesture string. For example, there will be limited ambiguity in the recognized gesture string when the majority of gestures are unambiguous deictic pointing gestures. Obviously, if more complex gestures are used, such as the multi-element gestures described in Johnston 1-3, other ways of combining the gesture and speech recognition systems may be more appropriate.
[0161] Accordingly, for the specific exemplary embodiment of the multimodal user input device 400 shown in FIG. 9, the gesture recognition system 200 first processes the incoming gestures to construct a gesture finite-state transducer, such as that shown in FIG. 22, corresponding to the range of gesture interpretations. In particular, in the systems and methods according to this invention, to abstract over specific content of the input gesture, the input gesture is represented as a finite state transducer. This finite-state transducer relates the gesture symbols as well as the specific contents, such as, for example, entities or points on the gesture input portion selected by the user, of the input gesture to the gesture symbols, including a symbol acting as a place holder for the specific contents of the input gesture. More particularly, one side of this gesture finite-state transducer contains the specific contents of the input gesture and the other side of the gesture finite-state transducer contains a symbol, for example, the “SEM” term, which acts as a place holder for the specific content of the input gesture which is on the other side of the gesture finite-state transducer. In addition, in order to perform multimodal integration using finite-state devices, the output projection of the gesture finite-state transducer, i.e., the side without the specific contents of the input gesture, is taken. For example, the output projection of the gesture finite-state transducer shown in FIG. 22 is shown as the gesture finite-state machine in FIG. 23. As discussed above, after multimodal integration, the gesture finite-state transducer is composed with a projection of the resulting gesture/speech/meaning finite-state machine in order to re-integrate the specific contents of the input gesture.
[0162] In addition, in the exemplary embodiments described above with respect to FIGS. 5 and 9, the gesture input is unambiguous. Thus, as shown in FIG. 23, a simple linearly-connected set of states forms the gesture finite-state machine. It should be appreciated that, if the received gestures involved more complex gesture recognition or were otherwise ambiguous, the recognized string of gestures would be represented as a lattice indicating all of the possible gesture recognitions and interpretations for the received gesture stream. Moreover, a weighted finite-state transducer could be used to incorporate the likelihoods of the various paths in such a lattice.
[0163] FIG. 24 is a flowchart outlining one exemplary embodiment of a method for combining and converting the various multimodal input streams into a combined finite-state transducer representing the semantic meaning of the combined multimodal input streams according to this invention. Beginning in step 500, control continues to step 510, where gesture and speech utterances are input through one or more input devices that together combine to form a multimodal user input device.
[0164] Next, in step 515, a gesture finite-state transducer is generated, from the input gesture utterance. The gesture finite-state transducer relates the specific contents, such as a list of entities, entity identifiers representing the entities selected by the user, or points on the gesture input portion selected by the user, of the gesture and the gesture symbols representing the input gesture to the gesture symbols representing the input gesture, including a symbol acting as a place holder for the specific contents of the input gesture.
[0165] This gesture finite-state transducer essentially stores the specific contents of the input gesture, such as entities or points on the gesture input portion, selected by the user, on one side of the gesture finite-state transducer and uses a symbol, such as, for example, the “SEM” term, as a place holder for the specific contents of the input gesture on the other side of gesture finite-state transducer.
[0166] Then, in step 520, a gesture finite-state machine is generated from the gesture finite-state transducer. The gesture finite-state machine is generated by taking a projection of the output of the gesture finite-state transducer. The output of the gesture finite-state transducer contains the gesture symbols including a symbol acting as a place holder for the specific contents of the input gesture. The projection of the output of the gesture finite state transducer, i.e. the gesture finite-state machine, can be used to perform a function, such as multimodal integration, using finite-state devices.
[0167] Alternatively, in step 515, a gesture lattice that represents the input gesture can be generated from the input gesture utterance. The gesture lattice contains gesture symbols and the specific contents of the gesture, such as entities or points selected by the user. In this case, in step 520, a gesture finite-state transducer is generated from the gesture finite-state machine and the gesture finite-state transducer relates the contents of the gesture lattice to gesture symbols representing the input gesture, including a symbol acting as a place holder for the specific contents of the input gesture.
[0168] Next, in step 530, the gesture finite-state machine is composed with the gesture-to-speech transducer to generate a gesture/language finite-state transducer. For example, in the exemplary embodiment described above, the gesture finite-state machine shown in FIG. 22 is composed with the gesture-to-speech finite-state transducer shown in FIG. 20 to form the gesture/language finite-state transducer shown in FIG. 25. The gesture/language finite-state transducer represents the relationship between the recognized stream of gestures and all of the possible word sequences that could occur with those gestures of the recognized stream of gestures.
[0169] Then, in step 540, in order to use this information to guide the speech recognition system 100, a projection of the gesture/language finite-state transducer is generated. In particular, a projection on the output tape or speech portion of the gesture/language finite-state transducer shown in FIG. 25 is taken to yield a finite-state machine.
[0170] Next, in step 550, the speech utterance is recognized using the projection of the gesture/language finite-state transducer as the language model. Using the projection of the gesture/language finite-state transducer as the language model enables the gestural information to directly influence the recognition process performed by the automatic speech recognition system 100. In particular, as shown in step 560, the automatic speech recognition system generates a word sequence lattice based on the projection of the gesture/language finite-state transducer in view of the word lattice 155. In the exemplary embodiment outlined above, using the projection of the gesture/language finite-state transducer shown in FIG. 25 as the language model for the speech recognition process results in the recognized word sequence lattice “phone for these three restaurants”, as shown in FIG. 26.
[0171] Then, in step 570, the gesture/language finite-state transducer is composed with the recognized word sequences lattice to generate a gesture/speech finite-state transducer. This reintegrates the gesture information that was removed when the projection of the gesture/language finite-state transducer was generated in step 540. The generated gesture/speech finite-state transducer contains the information both from the speech utterance and the gesture utterance received from the various portions of the multimodal user input device 400. For the example outlined above, composing the gesture/language finite-state transducer shown in FIG. 25 with the word sequences lattice shown in FIG. 26 generates the gesture/speech finite-state transducer shown in FIG. 27. Operation then continues to step 580.
[0172] Then, in step 580, the gesture/speech finite-state transducer is converted to a gesture/speech finite-state machine. In particular, the gesture/speech finite-state machine combines the input and output tapes of the gesture/speech finite-state transducer onto a single tape. In the exemplary embodiment outlined above, converting the gesture/speech finite-state transducer shown in FIG. 27 results in the gesture/speech finite-state machine shown in FIG. 28.
[0173] Next, in step 590, the gesture/speech finite-state machine is composed with the speech/gesture/meaning finite-state transducer shown in FIG. 21 to generate the meaning finite-state transducer shown in FIG. 29. Because the speech/gesture/meaning finite-state transducer relates the speech and gesture symbols to meaning, composing the gesture/speech finite-state machine results in the meaning finite-state transducer which captures the combined semantic meaning or representation contained in the independent modes input using the multimodal user input device. Thus, the meaning of the multimodal input received from the multimodal user input device can be read from the output tape of the meaning finite-state transducer. In the exemplary embodiment outlined above, composing the gesture/speech finite-state machine shown in FIG. 28 with the speech/gesture/meaning finite-state transducer shown in FIG. 21 results in the meaning finite-state transducer shown in FIG. 29. In particular, it should be appreciated that the meaning finite-state transducer shown in FIG. 29 is a linear finite-state transducer that unambiguously yields the meaning. Operation then continues to step 600.
[0174] In step 600, as shown in FIG. 30, a projection that factors out speech from the of the meaning finite-state transducer is taken and in step 610, as shown in FIG. 31, the projection is composed with the gesture input finite-state transducer in order to reincorporate the specific contents of the input gesture. Operation then continues to step 620. As shown in FIG. 32, for example, in step 620, a meaning is obtained.
[0175] It should be appreciated that, in embodiments that use much more complex multimodal interfaces, such as those illustrated in Johnston 1-3, the meaning finite-state transducer may very well be a weighted finite-state transducer having multiple paths between the start and end nodes representing the various possible meanings for the multimodal input and the probability corresponding to each path. In this case, in step 620, the most likely meaning would be selected from the meaning finite-state transducer based on the path through the meaning finite-state transducer having the highest probability. However, it should be appreciated that step 620 is optional and can be omitted. Then, in step 630, the process ends.
[0176] As outlined above, the various exemplary embodiments described herein allow spoken language and gesture input streams to be parsed and integrated by a single weighted finite-state device. This single weighted finite-state device provides language models for speech and gesture recognition and composes the meaning content from the speech and gesture input streams into a single semantic representation. Thus, the various systems and methods according to this invention not only address multimodal language recognition, but also encode the semantics as well as the syntax into a single weighted finite-state device. Compared to the previous approaches for integrating multimodal input streams, such as those described in Johnston 1-3, which compose elements from n-best lists of recognition results, the systems and methods according to this invention provide the potential for mutual compensation among the various multimodal input modes.
[0177] The systems and methods according to this invention allow the gestural input to dynamically alter the language model used for speech recognition. Additionally, the systems and methods according to this invention reduce the computational complexity of multi-dimensional multimodal parsing. In particular, the weighted finite-state devices used in the systems and methods according to this invention provide a well-understood probabilistic framework for combining the probability distributions associated with the speech and gesture input streams and for selecting among multiple competing multimodal interpretations.
[0178] It should be appreciated that the systems and methods for abstraction according to this invention, are not limited to multimodal integration. The systems and methods for abstraction according to this invention may be used in other functions implemented with finite-state devices in which it would be undesirable, difficult, impossible and/or unnecssary unnccessary to include certain components of the information represented by the finite-state devices when performing the desired function. In addition, it should be appreciated that the systems and methods for abstraction, according to this invention, may be used to abstract over certain components of the information represented by the finite-state devices that represent other forms of multimodal inputs.
[0179] It should be appreciated that the multimodal recognition and/or meaning system 1000 shown in FIG. 2, and/or each of the gesture recognition system 200, the multimodal parser/meaning recognition system 300 and/or the automatic speech recognition system 100 can each be implemented on a programmed general purpose computer. However, any or all of these systems can also be implemented on a special purpose computer, a programmed microprocessor or microcontroller and peripheral integrated circuit elements, an ASIC or other integrated circuit, a digital signal processor, a hardwired electronic or a logic circuit such as a discrete element circuit, a programmable logic device such as a PLD, a PLA, a FPGA or a PAL, or the like. In general, any device capable of implementing a finite-state machine that is in turn capable of implementing the flowchart shown in FIG. 22 and/or the various finite-state machines and transducers shown in FIGS. 19-21 and 23-28 can be used to implement one or more of the various systems shown in FIGS. 1-4.
[0180] Thus, it should be understood that each of the various systems and subsystems shown in FIGS. 1-4 can be implemented as portions of a suitably programmed general purpose computer. Alternatively, each of the systems or subsystems shown in FIGS. 1-4 can be implemented as physically distinct hardware circuits within an ASIC, or using a FPGA, a PLD, a PLA, or a PAL, or using discrete logic elements or discrete circuit elements. The particular form each of the systems and/or subsystems shown in FIGS. 1-4 will take is a design choice and will be obvious and predictable to those skilled in the art.
[0181] It should also be appreciated that, while the above-outlined description of the various systems and methods according to this invention and the figures focus on speech and gesture as the multimodal inputs, any known or later-developed set of two or more input streams representing different modes of information or communication, such as speech, electronic-ink-based gestures or other haptic modes, keyboard input, inputs generated by observing or sensing human body motions, including hand motions, gaze motions, facial expressions, or other human body motions, or any other known or later-developed method for communicating information, can be combined and used as one of the input streams in the multimodal utterance.
[0182] Thus, while this invention has been described in conjunction with the exemplary embodiments outlined above, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. Accordingly, the exemplary embodiments of these systems and methods according to this invention, as set forth above, are intended to be illustrative, not limiting. Various changes may be made without departing from the spirit and scope of this invention.
Claims
1. A method for combining related inputs of a single mode of a multimodal recognition device, comprising;
- receiving a plurality of inputs of the single mode of an input portion of the multimodal device;
- generating an input lattice from the plurality of inputs;
- combining at least two related inputs in the input lattice; and
- adding at least one branch to the input lattice, wherein the added branch represents the combination of the related inputs.
2. The method of claim 1, wherein:
- combining related inputs in the input lattice comprises combining related inputs in the input lattice; and
- the related inputs are temporally and/or spatially related.
3. The method of claim 1, wherein, for each of the at least two related inputs, receiving those inputs comprises receiving a selection, by a user, of at least one of an entity on the input portion, an area on the input portion and a location on the input portion.
4. The method of claim 3, wherein receiving the selection of the entity includes receiving a selection, by the user, of at least one of at least one object and at least one command on the input portion.
5. The method of claim 2, wherein combining related inputs comprises combining related inputs that represent at least two types of entities.
6. The method of claim 2, wherein: combining related inputs comprises combining related inputs that represent entities of a same type.
7. The method of claim 2, further comprising maintaining a record of the received plurality of inputs.
8. The method of claim 7, wherein maintaining the record of the received plurality of inputs comprises assigning at least one different weight to the received plurality of inputs than a weight assigned to the combined related inputs.
9. A method for combining adjacent inputs of a multimodal device comprising;
- receiving a plurality of inputs of one mode of the multimodal device in an input portion of the multimodal device;
- generating an input lattice from the plurality of inputs;
- combining at least two adjacent inputs on the input lattice; and
- adding at least one branch to the input lattice, wherein the added branch represents the combination of the adjacent inputs.
10. The method of claim 9, wherein combining adjacent inputs on the input lattice comprises combining adjacent inputs on the input lattice that are temporally related and/or spatially related.
11. The method of claim 10, wherein, for each of the at least two adjacent inputs, receiving those inputs comprises receiving a selection, by a user, of at least one of an entity on the input portion of the multimodal device, an area on the input portion of the multimodal device and a location on the input portion of the multimodal device.
12. The method of claim 11, wherein receiving the selection of the entity includes receiving a selection, by the user, of at least one of at least one object and at least one command on the input portion of the multimodal device.
13. The method of claim 11, wherein combining adjacent inputs comprises combining adjacent inputs that represent at least two types of entities.
14. The method of claim 11, wherein combining adjacent inputs comprises combining adjacent inputs that represent entities of a same type.
15. The method of claim 10, further comprising maintaining a record of the received plurality of inputs.
16. The method of claim 15, wherein maintaining the record of the received plurality of inputs comprises assigning at least one different weight to the received plurality of inputs than a weight assigned to the combined related inputs.
17. A method for combining related inputs, comprising:
- receiving a plurality of inputs in a first mode of a multimodal utterance;
- receiving at least one input in a second mode of a multimodal utterance, wherein the at least one input in the second mode of the multimodal utterance contains a reference to a plurality of items;
- generating an input lattice from the plurality of inputs in the first mode of the multimodal utterance;
- combining at least two related inputs in the first mode of the multimodal utterance; and
- adding at least one branch, representing at least one combination of related inputs, to the input lattice.
18. The method of claim 17, wherein combining related inputs in the first mode of the multimodal utterance comprises combining inputs in the input lattice that are temporally-related and/or spatially-related.
19. The method of claim 17, wherein combining the inputs in the first mode of the multimodal utterance comprises combining inputs in the first mode of the multimodal utterance that represent a same type of entity.
20. The method of claim 17, wherein combining the inputs in the first mode of the multimodal utterance comprises combining inputs in the first mode of the multimodal utterance that represent at least two types of entities.
21. The method of claim 17, wherein receiving at least one input in the second mode of the multimodal utterance comprises receiving an utterance in the second mode of the multimodal utterance that includes a reference to at least one of an entity, an area and a location on the input portion that are selected by a user.
22. The method of claim 17, further comprising adding at least one branch to the input lattice representing the plurality of inputs in the first mode of the multimodal utterance.
23. The method of claim 17, further comprising inputting an utterance in the second mode of the multimodal utterance, wherein the utterance includes at least one number representing a number of items to which the utterance in the second mode refers.
24. The method of claim 17, further comprising maintaining a record of the received plurality of inputs.
25. The method of claim 24, wherein maintaining the record of the received plurality of inputs comprises assigning at least one different weight to the received plurality of inputs than a weight assigned to the combined related inputs.
Type: Application
Filed: Aug 12, 2002
Publication Date: Mar 20, 2003
Applicant: AT&T Corp. (New York, NY)
Inventors: Michael J. Johnston (Hoboken, NJ), Srinivas Bangalore (Morristown, NJ)
Application Number: 10216448
