Matsushita Electric Industrial Co., Ltd.

Matsushita Electrical Industrial Co., Ltd.

Panasonic Technologies, Inc.

Patrick Nguyen

Jean-Claude Junqua

Luca Rigazio

Roland Kuhn

Robert C Boman

David Kryze

Philippe Gelin

Florent Perronnin

Younes Souilmi

Olivier Thyes

Ambroise Mutel

An improved method is provided for constructing compact acoustic models for use in a speech recognizer. The method includes: partitioning speech data from a plurality of training speakers according to at least one speech related criteria (i.e., vocal tract length); grouping together the partitioned speech data from training speakers having a similar speech characteristic; and training an acoustic bubble model for each group using the speech data within the group.

Bubble splitting for compact acoustic modeling

A media capture device has an audio input receptive of user speech relating to a media capture activity in close temporal relation to the media capture activity. A plurality of focused speech recognition lexica respectively relating to media capture activities are stored on the device, and a speech recognizer recognizes the user speech based on a selected one of the focused speech recognition lexica. A media tagger tags captured media with generated speech recognition text, and a media annotator annotates the captured media with a sample of the user speech that is suitable for input to a speech recognizer. Tagging and annotating are based on close temporal relation between receipt of the user speech and capture of the captured media. Annotations may be converted to tags during post processing, employed to edit a lexicon using letter-to-sound rules and spelled word input, or matched directly to speech to retrieve captured media.

Voice tagging, voice annotation, and speech recognition for portable devices with optional post processing

A method for performing noise adaptation of a target speech signal input to a speech recognition system, where the target speech signal contains both additive and convolutional noises. The method includes estimating an additive noise bias and a convolutional noise bias; in the target speech signal; and jointly compensating the target speech signal for the additive and convolutional noise biases in a feature domain.

Method and apparatus for feature domain joint channel and additive noise compensation

The dynamic programming technique employs a lexical tree that is encoded in computer memory as a flat representation in which the nodes of each generation occupy contiguous memory locations. The traversal algorithm employs a set of traversal rules whereby nodes of a given generation are processed before the parent nodes of that generation. The deepest child generation is processed first and traversal among nodes of each generation proceeds in the same topological direction.

Recognition system using lexical trees

The speech synthesizer is personalized to sound like or mimic the speech characteristics of an individual speaker. The individual speaker provides a quantity of enrollment data, which can be extracted from a short quantity of speech, and the system modifies the base synthesis parameters to more closely resemble those of the new speaker. More specifically, the synthesis parameters may be decomposed into speaker dependent parameters, such as context-independent parameters, and speaker independent parameters, such as context dependent parameters. The speaker dependent parameters are adapted using enrollment data from the new speaker. After adaptation, the speaker dependent parameters are combined with the speaker independent parameters to provide a set of personalized synthesis parameters. To adapt the parameters with a small amount of enrollment data, an eigenspace is constructed and used to constrain the position of the new speaker so that context independent parameters not provided by the new speaker may be estimated.

Voice personalization of speech synthesizer

Linear approximation of the background noise is applied after feature extraction and prior to speaker adaptation to allow the speaker adaptation system to adapt the speech models to the enrolling user without distortion from background noise. The linear approximation is applied in the feature domain, such as in the cepstral domain. Any adaptation technique that is commutative in the feature domain may be used.

Speaker and environment adaptation based on linear separation of variability sources

An embedded device for playing media files is capable of generating a play list of media files based on input speech from a user. It includes an indexer generating a plurality of speech recognition grammars. According to one aspect of the invention, the indexer generates speech recognition grammars based on contents of a media file header of the media file. According to another aspect of the invention, the indexer generates speech recognition grammars based on categories in a file path for retrieving the media file to a user location. When a speech recognizer receives an input speech from a user while in a selection mode, a media file selector compares the input speech received while in the selection mode to the plurality of speech recognition grammars, thereby selecting the media file.

System and method of media file access and retrieval using speech recognition

An e-mail message process is provided for use with a personal digital assistant which allows for the use of input speech messaging which is converted to text using a focused language model which is downloaded by a cellular phone connection to an Internet server which provides the focused language model based upon a topic for the intended e-mail message. The text that is generated from the input speech method can be summarized by the e-mail message processor and can be edited by the user. The generated e-mail message can then be transmitted again via cellular connection to an Internet e-mail server for transmitting the e-mail message to a recipient.

Focused language models for improved speech input of structured documents

A reduced dimensionality eigenvoice analytical technique is used during training to develop context-dependent acoustic models for allophones. Re-estimation processes are performed to more strongly separate speaker-dependent and speaker-independent components of the speech model. The eigenvoice technique is also used during run time upon the speech of a new speaker. The technique removes individual speaker idiosyncrasies, to produce more universally applicable and robust allophone models. In one embodiment the eigenvoice technique is used to identify the centroid of each speaker, which may then be 'subtracted out' of the recognition equation.

Eigenvoice re-estimation technique of acoustic models for speech recognition, speaker identification and speaker verification

A method is provided for improving pattern matching in a speech recognition system having a plurality of acoustic models. The improved method includes: receiving continuous speech input; generating a sequence of acoustic feature vectors that represent temporal and spectral behavior of the speech input; loading a first group of acoustic feature vectors from the sequence of acoustic feature vectors into a memory workspace accessible to a processor; loading an acoustic model from the plurality of acoustic models into the memory workspace; and determining a similarity measure for each acoustic feature vector of the first group of acoustic feature vectors in relation to the acoustic model. Prior to retrieving another group of acoustic feature vectors, similarity measures are computed for the first group of acoustic feature vectors in relation to each of the acoustic models employed by the speech recognition system. In this way, the improved method reduces the number I/O operations associated with loading and unloading each acoustic model into memory.

Pattern matching for large vocabulary speech recognition systems

Client speaker locations in a speaker space are used to generate speech models for comparison with test speaker data or test speaker speech models. The speaker space can be constructed using training speakers that are entirely separate from the population of client speakers, or from client speakers, or from a mix of training and client speakers. Reestimation of the speaker space based on client environment information is also provided to improve the likelihood that the client data will fall within the speaker space. During enrollment of the clients into the speaker space, additional client speech can be obtained when predetermined conditions are met. The speaker distribution can also be used in the client enrollment step.

Speaker verification and speaker identification based on a priori knowledge

Methods and apparatus for blind channel estimation of a speech signal corrupted by a communication channel are provided. One method includes converting a noisy speech signal into either a cepstral representation or a log-spectral representation; estimating a correlation of the representation of the noisy speech signal; determining an average of the noisy speech signal; constructing and solving, subject to a minimization constraint, a system of linear equations utilizing a correlation structure of a clean speech training signal, the correlation of the representation of the noisy speech signal, and the average of the noisy speech signal; and selecting a sign of the solution of the system of linear equations to estimate an average clean speech signal in a processing window.

Methods and apparatus for blind channel estimation based upon speech correlation structure

A set of speaker dependent models or adapted models is trained upon a comparatively large number of training speakers, one model per speaker, and model parameters are extracted in a predefined order to construct a set of supervectors, one per speaker. Dimensionality reduction is then performed on the set of supervectors to generate a set of eigenvectors that define an eigenvoice space. If desired, the number of vectors may be reduced to achieve data compression. Thereafter, a new speaker provides adaptation data from which a supervector is constructed by constraining this supervector to be in the eigenvoice space based on a maximum likelihood estimation. The resulting coefficients in the eigenspace of this new speaker may then be used to construct a new set of model parameters from which an adapted model is constructed for that speaker. The adapted model may then be further adapted via MAP, MLLR, MLED or the like. The eigenvoice technique may be applied to MLLR transformation matrices or the like; Bayesian estimation performed in eigenspace uses prior knowledge about speaker space density to refine the estimate about the location of a new speaker in eigenspace.

Dimensionality reduction for speaker normalization and speaker and environment adaptation using eigenvoice techniques

Supervised adaptation speech is supplied to the recognizer and the recognizer generates the N-best transcriptions of the adaptation speech. These transcriptions include the one transcription known to be correct, based on a priori knowledge of the adaptation speech, and the remaining transcriptions known to be incorrect. The system applies weights to each transcription: a positive weight to the correct transcription and negative weights to the incorrect transcriptions. These weights have the effect of moving the incorrect transcriptions away from the correct one, rendering the recognition system more discriminative for the new speaker's speaking characteristics. Weights applied to the incorrect solutions are based on the respective likelihood scores generated by the recognizer. The sum of all weights (positive and negative) are a positive number. This ensures that the system will converge.

Supervised adaptation using corrective N-best decoding

A set of speaker dependent models is trained upon a comparatively large number of training speakers, one model per speaker, and model parameters are extracted in a predefined order to construct a set of supervectors, one per speaker. Principle component analysis is then performed on the set of supervectors to generate a set of eigenvectors that define an eigenvoice space. If desired, the number of vectors may be reduced to achieve data compression. Thereafter, a new speaker provides adaptation data from which a supervector is constructed by constraining this supervector to be in the eigenvoice space based on a maximum likelihood estimation. The resulting coefficients in the eigenspace of this new speaker may then be used to construct a new set of model parameters from which an adapted model is constructed for that speaker. Environmental adaptation may be performed by including environmental variations in the training data.

Growing community of inventors

Santa Barbara, CA, United States of America

Patrick Nguyen