Abstract: Decision trees are used to store a series of yes-no questions that can be used to convert spelled-word letter sequences into pronunciations. Letter-only trees, having internal nodes populated with questions about letters in the input sequence, generate one or more pronunciations based on probability data stored in the leaf nodes of the tree. The pronunciations may then be improved by processing them using mixed trees which are populated with questions about letters in the sequence and also questions about phonemes associated with those letters. The mixed tree screens out pronunciations that would not occur in natural speech, thereby greatly improving the results of the letter-to-pronunciation transformation.

Abstract: The system performs unsupervised speech model adaptation using the recognizer to generate the N-best solutions for an input utterance. Each of these N-best solutions is tested by a reliable information extraction process. Reliable information is extracted by a weighting technique based on likelihood scores generated by the recognizer, or by a non-linear thresholding function. The system may be used in a single pass implementation or iteratively in a multi-pass implementation.

Abstract: The speech recognizer incorporates a language model that reduces the number of acoustic pattern matching sequences that must be performed by the recognizer. The language model is based on knowledge of a pre-defined set of syntactically defined content and includes a data structure that organizes the content according to acoustic confusability. A spelled name recognition system based on the recognizer employs a language model based on classes of letters that the recognizer frequently confuses for one another. The language model data structure is optionally an N-gram data structure, a tree data structure, or an incrementally configured network that is built during a training sequence. The incrementally configured network has nodes that are selected based on acoustic distance from a predetermined lexicon.

Abstract: Speech models are constructed and trained upon the speech of known client speakers (and also impostor speakers, in the case of speaker verification). Parameters from these models are concatenated to define supervectors and a linear transformation upon these supervectors results in a dimensionality reduction yielding a low-dimensional space called eigenspace. The training speakers are then represented as points or distributions in eigenspace. Thereafter, new speech data from the test speaker is placed into eigenspace through a similar linear transformation and the proximity in eigenspace of the test speaker to the training speakers serves to authenticate or identify the test speaker.

Abstract: A two-stage pronunciation generator utilizes mixed decision trees that includes a network of yes-no questions about letter, syntax, context, and dialect in a spelled word sequence. A second stage utilizes decision trees that includes a network of yes-no questions about adjacent phonemes in the phoneme sequence corresponding to the spelled word sequence. Leaf nodes of the mixed decision trees provide information about which phonetic transcriptions are most probable. Using the mixed trees, scores are developed for each of a plurality of possible pronunciations, and these scores can be used to select the best pronunciation as well as to rank pronunciations in order of probability. The pronunciations generated by the system can be used in speech synthesis and speech recognition applications as well as lexicography applications.

Abstract: The mixed decision tree includes a network of yes-no questions about adjacent letters in a spelled word sequence and also about adjacent phonemes in the phoneme sequence corresponding to the spelled word sequence. Leaf nodes of the mixed decision tree provide information about which phonetic transcriptions are most probable. Using the mixed trees, scores are developed for each of a plurality of possible pronunciations, and these scores can be used to select the best pronunciation as well as to rank pronunciations in order of probability. The pronunciations generated by the system can be used in speech synthesis and speech recognition applications as well as lexicography applications.

Abstract: The input speech is segmented using plural grammar networks, including a network that includes a filler model designed to represent noise or extraneous speech. Recognition processing results in plural lists of candidates, each list containing the N-best candidates generated. The lists are then separately aligned with the dictionary of valid names to generate two lists of valid names. The final recognition pass combines these two lists of names into a dynamic grammar and this dynamic grammar may be used to find the best candidate name using Viterbi recognition. A telephone call routing application based on the recognition system selects the best candidate name corresponding to the name spelled by the user, whether the user pronounces the name prior to spelling, or not.

Abstract: The voice dialing server plugs into one or more unused extensions of a branch exchange system to provide each of the users on the system with voice dialing services. To use the system a user simply dials the extension to which the server is attached. The server then prompts the user to supply the name of a party to be called. The name is then looked up in a telephone number dictionary unique to that user. The system then places the telephone call by sending commands to the branch exchange system that simulate the operations a user would perform to connect to an outside line or inside extension and then place the call. The server incorporates a speech processing module having a multistage word recognizer that represents speech in terms of high phoneme similarity values. This representation is highly compact, allowing the word recognizer to perform the recognizer and fine match stages with far less processor overhead than frame-by-frame speech recognizers.

Abstract: The multimodal telephone prompts the user using both a visual display and synthesized voice. It receives user input via keypad and programmable soft keys associated with the display, and also through user-spoken commands. The voice module includes a two stage speech recognizer that models speech in terms of high similarity values. A dialog manager associated with the voice module maintains the visual and verbal systems in synchronism with one another. The dialog manager administers a state machine that records the dialog context. The dialog context is used to ensure that the appropriate visual prompts are displayed--showing what commands are possible at any given point in the dialog. The speech recognizer also uses the dialog context to select the recognized word candidate that is appropriate to the current context.

Abstract: The clustering technique produces a low complexity and yet high accuracy speech representation for use with speech recognizers. The task database comprising the test speech to be modeled is segmented into subword units such as phonemes and labeled to indicate each phoneme in its left and right context (triphones). Hidden Markov Models are constructed for each context-independent phoneme and trained. Then the center states are tied for all phonemes of the same class. Triphones are trained and all poorly-trained models are eliminated by merging their training data with the nearest well-trained model using a weighted divergence computation to ascertain distance. Before merging, the threshold for each class is adjusted until the number of good models for each phoneme class is within predetermined upper and lower limits. Finally, if desired, the number of mixture components used to represent each model may be increased and the models retrained. This latter step increases the accuracy.

Abstract: The call routing device plugs into existing extensions of the office telephone network or PBX system and acts as a "virtual" operator, prompting incoming callers to spell the name of the desired recipient. The speech recognizer uses a multipass procedure employing Hidden Markov Models and dynamic programming. The N-best hypotheses are propagated between passes, allowing the more computationally costly routines to be reserved until the final pass, when the size of the search space is significantly reduced. The routing device prompts the user to confirm that the selected name is correct, whereupon the device signals the telephone network to automatically switch the incoming call to the telephone extension of the selected recipient.

Abstract: The language acquisition system assists the user in acquiring the language of an application. The system uses the dialogue context, a dialogue model and syntactic-semantic grammars to progressively build commands which, to the application program, are syntactically and semantically correct in the current context and which can be interpreted by the dialogue server which then controls the application program. The system is independent of any particular application language. The system is also multimodal and supports both speech and text input. A toolkit is provided to add this functionality to virtually any application program.

Abstract: A multipass recognition strategy selects the N-best hypotheses resulting from each pass and propagates these N-best to the next pass. This strategy outperforms conventional hidden Markov model recognizers using a grammar constraining all possible names. Real time recognition of continuously spelled names is made feasible, in part, because the processor-intensive costly constraints are applied, if at all, in the 4th pass, after the system has produced a much smaller dynamic grammar.

Abstract: The dictionary is broken into clusters by first grouping the dictionary according to a rule based procedure whereby the dictionary is sorted by word length and alphabetically. After sorting, a plurality of first cluster centers is generated by selecting the dictionary entries that differ from neighboring entries by the first letter. Each of the dictionary entries is then assigned to the closest one of the first cluster centers using a dynamic time warping procedure. These newly formed clusters are then each analyzed to find the true cluster center and the dictionary entries are then each assigned to the closest true cluster center. The clusters, so formed, may then be rapidly searched to locate any dictionary entry. The search is quite efficient because only the closest cluster to the desired dictionary entry needs to be searched.

Abstract: A method for analyzing a speech signal to isolate speech and nonspeech portions of the speech signal is provided. The method is applied to an input speech signal to determine boundary values locating isolated words or groups of words within the speech signal. First, a comparison signal is generated which is biased to emphasize components of the signal having preselected frequencies. Next, the system compares the comparison signal with a threshold level to determine estimated boundary values demonstrating the beginning and ending points of the words. Once the estimated boundary values are calculated, the system adjusts the boundary values to achieve final boundary values. The specific amount of adjustment varies, depending upon the amount of noise present in the signal. The final pair of boundary values provide a reliable indication of the location and duration of the isolated word or group of words within the speech signal.