Abstract: Disclosed herein are systems, computer-implemented methods, and tangible computer-readable media for handling missing speech data. The computer-implemented method includes receiving speech with a missing segment, generating a plurality of hypotheses for the missing segment, identifying a best hypothesis for the missing segment, and recognizing the received speech by inserting the identified best hypothesis for the missing segment. In another method embodiment, the final step is replaced with synthesizing the received speech by inserting the identified best hypothesis for the missing segment. In one aspect, the method further includes identifying a duration for the missing segment and generating the plurality of hypotheses of the identified duration for the missing segment. The step of identifying the best hypothesis for the missing segment can be based on speech context, a pronouncing lexicon, and/or a language model. Each hypothesis can have an identical acoustic score.

Abstract: Disclosed herein are systems, methods, and non-transitory computer-readable storage media for generating a synthetic voice. A system configured to practice the method combines a first database of a first text-to-speech voice and a second database of a second text-to-speech voice to generate a combined database, selects from the combined database, based on a policy, voice units of a phonetic category for the synthetic voice to yield selected voice units, and synthesizes speech based on the selected voice units. The system can synthesize speech without parameterizing the first text-to-speech voice and the second text-to-speech voice. A policy can define, for a particular phonetic category, from which text-to-speech voice to select voice units. The combined database can include multiple text-to-speech voices from different speakers. The combined database can include voices of a single speaker speaking in different styles. The combined database can include voices of different languages.

Abstract: Disclosed herein are systems, computer-implemented methods, and computer-readable storage media for recognizing speech by adapting automatic speech recognition pronunciation by acoustic model restructuring. The method identifies an acoustic model and a matching pronouncing dictionary trained on typical native speech in a target dialect. The method collects speech from a new speaker resulting in collected speech and transcribes the collected speech to generate a lattice of plausible phonemes. Then the method creates a custom speech model for representing each phoneme used in the pronouncing dictionary by a weighted sum of acoustic models for all the plausible phonemes, wherein the pronouncing dictionary does not change, but the model of the acoustic space for each phoneme in the dictionary becomes a weighted sum of the acoustic models of phonemes of the typical native speech. Finally the method includes recognizing via a processor additional speech from the target speaker using the custom speech model.

Abstract: Disclosed herein are systems, methods, and non-transitory computer-readable storage media for detecting and correcting abnormal stress patterns in unit-selection speech synthesis. A system practicing the method detects incorrect stress patterns in selected acoustic units representing speech to be synthesized, and corrects the incorrect stress patterns in the selected acoustic units to yield corrected stress patterns. The system can further synthesize speech based on the corrected stress patterns. In one aspect, the system also classifies the incorrect stress patterns using a machine learning algorithm such as a classification and regression tree, adaptive boosting, support vector machine, and maximum entropy. In this way a text-to-speech unit selection speech synthesizer can produce more natural sounding speech with suitable stress patterns regardless of the stress of units in a unit selection database.

Abstract: A system, method and computer-readable storage devices are for normalizing text for ASR and TTS in a language-neutral way. The system described herein divides Unicode text into meaningful chunks called “atomic tokens.” The atomic tokens strongly correlate to their actual pronunciation, and not to their meaning The system combines the tokenization with a data-driven classification scheme, followed by class-determined actions to convert text to normalized form. The classification labels are based on pronunciation, unlike alternative approaches that typically employ Named Entity-based categories. Thus, this approach is relatively simple to adapt to new languages. Non-experts can easily annotate training data because the tokens are based on pronunciation alone.

Abstract: Disclosed herein are systems, methods, and non-transitory computer-readable storage media for reducing latency in web-browsing TTS systems without the use of a plug-in or Flash® module. A system configured according to the disclosed methods allows the browser to send prosodically meaningful sections of text to a web server. A TTS server then converts intonational phrases of the text into audio and responds to the browser with the audio file. The system saves the audio file in a cache, with the file indexed by a unique identifier. As the system continues converting text into speech, when identical text appears the system uses the cached audio corresponding to the identical text without the need for re-synthesis via the TTS server.

Abstract: Disclosed herein are systems, computer-implemented methods, and computer-readable storage media for recognizing speech by adapting automatic speech recognition pronunciation by acoustic model restructuring. The method identifies an acoustic model and a matching pronouncing dictionary trained on typical native speech in a target dialect. The method collects speech from a new speaker resulting in collected speech and transcribes the collected speech to generate a lattice of plausible phonemes. Then the method creates a custom speech model for representing each phoneme used in the pronouncing dictionary by a weighted sum of acoustic models for all the plausible phonemes, wherein the pronouncing dictionary does not change, but the model of the acoustic space for each phoneme in the dictionary becomes a weighted sum of the acoustic models of phonemes of the typical native speech. Finally the method includes recognizing via a processor additional speech from the target speaker using the custom speech model.

Abstract: Disclosed herein are systems, computer-implemented methods, and tangible computer-readable media for handling missing speech data. The computer-implemented method includes receiving speech with a missing segment, generating a plurality of hypotheses for the missing segment, identifying a best hypothesis for the missing segment, and recognizing the received speech by inserting the identified best hypothesis for the missing segment. In another method embodiment, the final step is replaced with synthesizing the received speech by inserting the identified best hypothesis for the missing segment. In one aspect, the method further includes identifying a duration for the missing segment and generating the plurality of hypotheses of the identified duration for the missing segment. The step of identifying the best hypothesis for the missing segment can be based on speech context, a pronouncing lexicon, and/or a language model. Each hypothesis can have an identical acoustic score.

Abstract: Systems, methods, and computer-readable storage media for intelligent caching of concatenative speech units for use in speech synthesis. A system configured to practice the method can identify a speech synthesis context, and determine, based on a local cache of text-to-speech units for a text-to-speech voice and based on the speech synthesis context, additional text-to-speech units which are not in the local cache. The system can request from a server the additional text-to-speech units, and store the additional text-to-speech units in the local cache. The system can then synthesize speech using the text-to-speech units and the additional text-to-speech units in the local cache. The system can prune the cache as the context changes, based on availability of local storage, or after synthesizing the speech. The local cache can store a core set of text-to-speech units associated with the text-to-speech voice that cannot be pruned from the local cache.

Abstract: Disclosed herein are systems, methods, and non-transitory computer-readable storage media for generating a synthetic voice. A system configured to practice the method combines a first database of a first text-to-speech voice and a second database of a second text-to-speech voice to generate a combined database, selects from the combined database, based on a policy, voice units of a phonetic category for the synthetic voice to yield selected voice units, and synthesizes speech based on the selected voice units. The system can synthesize speech without parameterizing the first text-to-speech voice and the second text-to-speech voice. A policy can define, for a particular phonetic category, from which text-to-speech voice to select voice units. The combined database can include multiple text-to-speech voices from different speakers. The combined database can include voices of a single speaker speaking in different styles. The combined database can include voices of different languages.

Abstract: Disclosed herein are systems, methods, and non-transitory computer-readable storage media for detecting and correcting abnormal stress patterns in unit-selection speech synthesis. A system practicing the method detects incorrect stress patterns in selected acoustic units representing speech to be synthesized, and corrects the incorrect stress patterns in the selected acoustic units to yield corrected stress patterns. The system can further synthesize speech based on the corrected stress patterns. In one aspect, the system also classifies the incorrect stress patterns using a machine learning algorithm such as a classification and regression tree, adaptive boosting, support vector machine, and maximum entropy. In this way a text-to-speech unit selection speech synthesizer can produce more natural sounding speech with suitable stress patterns regardless of the stress of units in a unit selection database.

Abstract: A system, method and computer-readable storage devices are for using a single set of normalization protocols and a single language lexica (or dictionary) for both TTS and ASR. The system receives input (which is either text to be converted to speech or ASR training text), then normalizes the input. The system produces, using the normalized input and a dictionary configured for both automatic speech recognition and text-to-speech processing, output which is either phonemes corresponding to the input or text corresponding to the input for training the ASR system. When the output is phonemes corresponding to the input, the system generates speech by performing prosody generation and unit selection synthesis using the phonemes. When the output is text corresponding to the input, the system trains both an acoustic model and a language model for use in future speech recognition.

Abstract: Disclosed herein are systems, methods, and non-transitory computer-readable storage media for reducing latency in web-browsing TTS systems without the use of a plug-in or Flash® module. A system configured according to the disclosed methods allows the browser to send prosodically meaningful sections of text to a web server. A TTS server then converts intonational phrases of the text into audio and responds to the browser with the audio file. The system saves the audio file in a cache, with the file indexed by a unique identifier. As the system continues converting text into speech, when identical text appears the system uses the cached audio corresponding to the identical text without the need for re-synthesis via the TTS server.

Abstract: Systems, methods, and computer-readable storage media for intelligent caching of concatenative speech units for use in speech synthesis. A system configured to practice the method can identify a speech synthesis context, and determine, based on a local cache of text-to-speech units for a text-to-speech voice and based on the speech synthesis context, additional text-to-speech units which are not in the local cache. The system can request from a server the additional text-to-speech units, and store the additional text-to-speech units in the local cache. The system can then synthesize speech using the text-to-speech units and the additional text-to-speech units in the local cache. The system can prune the cache as the context changes, based on availability of local storage, or after synthesizing the speech. The local cache can store a core set of text-to-speech units associated with the text-to-speech voice that cannot be pruned from the local cache.

Abstract: A system, method and computer readable medium that enhances a speech database for speech synthesis is disclosed. The method may include labeling audio files in a primary speech database, identifying segments in the labeled audio files that have varying pronunciations based on language differences, identifying replacement segments in a secondary speech database, enhancing the primary speech database by substituting the identified secondary speech database segments for the corresponding identified segments in the primary speech database, and storing the enhanced primary speech database for use in speech synthesis.

Abstract: Systems, methods, and computer-readable storage devices for generating speech using a presentation style specific to a user, and in particular the user's social group. Systems configured according to this disclosure can then use the resulting, personalized, text and/or speech in a spoken dialogue or presentation system to communicate with the user. For example, a system practicing the disclosed method can receive speech from a user, identify the user, and respond to the received speech by applying a personalized natural language generation model. The personalized natural language generation model provides communications which can be specific to the identified user.

Abstract: Systems, methods, and computer-readable storage devices to improve the quality of synthetic speech generation. A system selects speech units from a speech unit database, the speech units corresponding to text to be converted to speech. The system identifies a desired prosodic curve of speech produced from the selected speech units, and also identifies an actual prosodic curve of the speech units. The selected speech units are modified such that a new prosodic curve of the modified speech units matches the desired prosodic curve. The system stores the modified speech units into the speech unit database for use in generating future speech, thereby increasing the prosodic coverage of the database with the expectation of improving the output quality.

Abstract: Disclosed herein are methods for presenting speech from a selected text that is on a computing device. This method includes presenting text on a touch-sensitive display and having that text size within a threshold level so that the computing device can accurately determine the intent of the user when the user touches the touch screen. Once the user touch has been received, the computing device identifies and interprets the portion of text that is to be selected, and subsequently presents the text audibly to the user.

Abstract: Disclosed herein are methods, systems, and computer-readable storage media for automatic speech recognition. The method includes selecting a speaker independent model, and selecting a quantity of speaker dependent models, the quantity of speaker dependent models being based on available computing resources, the selected models including the speaker independent model and the quantity of speaker dependent models. The method also includes recognizing an utterance using each of the selected models in parallel, and selecting a dominant speech model from the selected models based on recognition accuracy using the group of selected models. The system includes a processor and modules configured to control the processor to perform the method. The computer-readable storage medium includes instructions for causing a computing device to perform the steps of the method.

Abstract: Disclosed herein are systems, computer-implemented methods, and computer-readable storage media for recognizing speech by adapting automatic speech recognition pronunciation by acoustic model restructuring. The method identifies an acoustic model and a matching pronouncing dictionary trained on typical native speech in a target dialect. The method collects speech from a new speaker resulting in collected speech and transcribes the collected speech to generate a lattice of plausible phonemes. Then the method creates a custom speech model for representing each phoneme used in the pronouncing dictionary by a weighted sum of acoustic models for all the plausible phonemes, wherein the pronouncing dictionary does not change, but the model of the acoustic space for each phoneme in the dictionary becomes a weighted sum of the acoustic models of phonemes of the typical native speech. Finally the method includes recognizing via a processor additional speech from the target speaker using the custom speech model.