Abstract: The invention concerns a method and system for detecting morphemes in a user's communication. The method may include recognizing a lattice of phone strings from the user's input communication, the lattice representing a distribution over the phone strings, and detecting morphemes in the user's input communication using the lattice. The morphemes may be acoustic and/or non-acoustic. The morphemes may represent any unit or sub-unit of communication including phones, diphones, phone-phrases, syllables, grammars, words, gestures, tablet strokes, body movements, mouse clicks, etc. The training speech may be verbal, non-verbal, a combination of verbal and non-verbal, or multimodal.

Abstract: Systems and methods for annotating speech data. The present invention reduces the time required to annotate speech data by selecting utterances for annotation that will be of greatest benefit. A selection module uses speech models, including speech recognition models and spoken language understanding models, to identify utterances that should be annotated based on criteria such as confidence scores generated by the models. These utterances are placed in an annotation list along with a type of annotation to be performed for the utterances and an order in which the annotation should proceed. The utterances in the annotation list can be annotated for speech recognition purposes, spoken language understanding purposes, labeling purposes, etc. The selection module can also select utterances for annotation based on previously annotated speech data and deficiencies in the various models.

Abstract: The invention concerns a method and system for detecting morphemes in a user's communication. The method may include recognizing a lattice of phone strings from the user's input communication, the lattice representing a distribution over the phone strings, and detecting morphemes in the user's input communication using the lattice. The morphemes may be acoustic and/or non-acoustic. The morphemes may represent any unit or sub-unit of communication including phones, diphones, phone-phrases, syllables, grammars, words, gestures, tablet strokes, body movements, mouse clicks, etc. The training speech may be verbal, non-verbal, a combination of verbal and non-verbal, or multimodal.

Abstract: A method is disclosed for designing a labeling guide for use by a labeler in labeling data used for training a spoken language understanding (SLU) module for an application. The method comprises a labeling guide designer selecting domain-independent actions applicable to an application, selecting domain-dependent objects according to characteristics of the application, and generating a labeling guide using the selected domain-independent actions and selected domain-dependent objects. An advantage of the labeling guide generated in this manner is that the labeling guide designer can easily port the labeling guide to a new application by selecting a set of domain-independent action and then selecting the domain-dependent objects related to the new application.

Abstract: A method of grammar learning from a corpus comprises, for the other non-context words, generating frequency vectors for each non-context token in a corpus based upon counted occurrences of a predetermined relationship of the non-context tokens to identified context tokens. Clusters are grown from the frequency vectors according to a lexical correlation among the non-context tokens.

Abstract: The invention concerns a method and system for detecting morphemes in a user's communication. The method may include recognizing a lattice of phone strings from the user's input communication, the lattice representing a distribution over the phone strings, and detecting morphemes in the user's input communication using the lattice. The morphemes may be acoustic and/or non-acoustic. The morphemes may represent any unit or sub-unit of communication including phones, diphones, phone-phrases, syllables, grammars, words, gestures, tablet strokes, body movements, mouse clicks, etc. The training speech may be verbal, non-verbal, a combination of verbal and non-verbal, or multimodal.

Abstract: A large amount of human labor is required to transcribe and annotate a training corpus that is needed to create and update models for automatic speech recognition (ASR) and spoken language understanding (SLU). Active learning enables a reduction in the amount of transcribed and annotated data required to train ASR and SLU models. In one aspect of the present invention, an active learning ASR process and active learning SLU process are coupled, thereby enabling further efficiencies to be gained relative to a process that maintains an isolation of data in both the ASR and SLU domains.

Abstract: The invention concerns a method and system for detecting morphemes in a user's communication. The method may include recognizing a lattice of phone strings from the user's input communication, the lattice representing a distribution over the phone strings, and detecting morphemes in the user's input communication using the lattice. The morphemes may be acoustic and/or non-acoustic. The morphemes may represent any unit or sub-unit of communication including phones, diphones, phone-phrases, syllables, grammars, words, gestures, tablet strokes, body movements, mouse clicks, etc. The training speech may be verbal, non-verbal, a combination of verbal and non-verbal, or multimodal.

Abstract: A method, system and machine-readable medium are provided for watermarking natural language digital text. A deep structure may be generated and a group of features may be extracted from natural language digital text input. The deep structure may be modified based, at least partly, on a watermark. Natural language digital text output may be generated based on the modified deep structure.

Abstract: Disclosed is a system and method of decomposing a lattice transition matrix into a block diagonal matrix. The method is applicable to automatic speech recognition but can be used in other contexts as well, such as parsing, named entity extraction and any other methods. The method normalizes the topology of any input graph according to a canonical form.

Abstract: A voice-enabled help desk service is disclosed. The service comprises an automatic speech recognition module for recognizing speech from a user, a spoken language understanding module for understanding the output from the automatic speech recognition module, a dialog management module for generating a response to speech from the user, a natural voices text-to-speech synthesis module for synthesizing speech to generate the response to the user, and a frequently asked questions module. The frequently asked questions module handles frequently asked questions from the user by changing voices and providing predetermined prompts to answer the frequently asked question.

Abstract: A system for recognizing connected digits in natural spoken dialogue includes a speech recognition processor that receives unconstrained fluent input speech and produces a string of words that can include a numeric language, and a numeric understanding processor that converts the string of words into a sequence of digits based on a set of rules. An acoustic model database utilized by the speech recognition processor includes a first set of hidden Markov models that characterize the acoustic features of numeric words and phrases, a second set of hidden Markov models that characterize the acoustic features of the remaining vocabulary words, and a filler model that characterizes the acoustic features of out-of-vocabulary utterances. An utterance verification processor verifies the accuracy of the string of words. A validation database stores a grammar, and a string validation processor outputs validity information based on a comparison of the sequence of digits with the grammar.

Abstract: State-of-the-art speech recognition systems are trained using transcribed utterances, preparation of which is labor-intensive and time-consuming. The present invention is an iterative method for reducing the transcription effort for training in automatic speech recognition (ASR). Active learning aims at reducing the number of training examples to be labeled by automatically processing the unlabeled examples and then selecting the most informative ones with respect to a given cost function for a human to label. The method comprises automatically estimating a confidence score for each word of the utterance and exploiting the lattice output of a speech recognizer, which was trained on a small set of transcribed data. An utterance confidence score is computed based on these word confidence scores; then the utterances are selectively sampled to be transcribed using the utterance confidence scores.

Abstract: The invention concerns a method of generating morphemes for speech recognition and understanding. The method may include receiving training speech, selecting candidate sub-morphemes from the training speech, selecting salient sub-morphemes from the candidate sub-morphemes based on salience measurements, and clustering the salient sub-morphemes based on semantic and syntactic similarities into morphemes. The morphemes may be acoustic and/or non-acoustic. The sub-morphemes may represent any sub-unit of communication including phones, phone-phrases, grammars, diphones, words, gestures, tablet strokes, body movements, mouse clicks, etc. The training speech may be verbal, non-verbal, a combination of verbal and non-verbal, or multimodal.

Abstract: A method and apparatus for stochastic finite-state machine translation is provided. The method may include receiving a speech input and translating the speech input in a source language into one or more symbols in a target language based on stochastic language model. Subsequently, all possible sequences of the translated symbols may be generated. One of the generated sequences may be selected based on a monolingual target language model.

Abstract: Utterance data that includes at least a small amount of manually transcribed data is provided. Automatic speech recognition is performed on ones of the utterance data not having a corresponding manual transcription to produce automatically transcribed utterances. A model is trained using all of the manually transcribed data and the automatically transcribed utterances. A predetermined number of utterances not having a corresponding manual transcription are intelligently selected and manually transcribed. Ones of the automatically transcribed data as well as ones having a corresponding manual transcription are labeled. In another aspect of the invention, audio data is mined from at least one source, and a language model is trained for call classification from the mined audio data to produce a language model.

Abstract: Systems and methods for determining word confidence scores. Speech recognition systems generate a word lattice for speech input. Posterior probabilities of the words in the word lattice are determined using a forward-backward algorithm. Next, time slots are defined for the word lattice, and for all transitions that at least partially overlap a particular time slot, the posterior probabilities of transitions that have the same word label are combined for those transitions. The combined posterior probabilities are used as confidence scores. A local entropy can be computed on the competitor transitions of a particular time slot and also used as a confidence score.

Abstract: The invention concerns a method of task classification using morphemes which operates on the task objective of a user. The morphemes may be generated by clustering selected ones of the salient sub-morphemes selected from training speech which are semantically and syntactically similar. The method may include detecting morphemes present in the user's input communication, and making task-type classification decisions based on the detected morphemes in the user's input communication. The morphemes may be verbal and/or non-verbal.

Abstract: A spoken dialog system configured to use a combined confidence score. A first confidence score, indicating a confidence level in a speech recognition result of recognizing an utterance, is provided. A second confidence level, indicating a confidence level of mapping the speech recognition result to an intent, is provided. The first confidence score and the second confidence score are combined to form a combined confidence score. A determination is made, with respect to whether to accept the intent, based on the combined confidence score.

Abstract: In a method of learning grammar from a corpus, context words are identified from a corpus. For the other non-context words, the method counts the occurrence of predetermined relationships which the context words, and maps the counted occurrences to a multidimensional frequency space. Clusters are grown from the frequency vectors. The clusters represent classes of words; words in the same cluster possess the same lexical significancy and provide an indicator of grammatical structure.