International Business Machines Corporation

Nuance Communications, Inc.

Peder Andreas Olsen

Vaibhava Goel

Steven John Rennie

Charles A Micchelli

Michael Alan Picheny

Jing Chang Huang

Harry William Printz

Ramesh Ambat Gopinath

Sankar Basu

John Randall Hershey

Antonio R Lee

Petr Novak

Dimitri Kanevsky

Liam David Comerford

Karthik Visweswariah

Etienne Marcheret

Ellen Marie Eide

Chengjun Julian Chen

Ramesh Natarajan

Amit Dhurandhar

Karthikeyan Shanmugam

Martin Franz

Neal J Alewine

Michael Daniel Monkowski

Peter V De Souza

Sholom M Weiss

Satyanarayana Dharanipragada

Pierre L Dognin

Petr Fousek

John W Eckhart

Ronny Luss

Kenneth D White

Catalina M Danis

Scott Elliot Axelrod

Satya Dharanipragada

Trausti Thor Kristjansson

A method, system, and computer program product, including generating, using a linear probe, confidence scores through flattened intermediate representations and theoretically-justified weighting of samples during a training of the simple model using the confidence scores of the intermediate representations.

Simple models using confidence profiles

Software that performs the following steps: (i) receiving data from a first database and data from a second database, (ii) identifying a training subset and a test subset from the received data; (iii) generating a first graphical model using data from the training subset; (iv) generating a second graphical model using data from the training subset; (v) determining respective weights for the first graphical model and the second graphical model by using an expectation maximization method on data from the test subset; (vi) generating a third graphical model by interpolating at least the first graphical model and the second graphical model using their respectively determined weights; and (vii) defining one or more links between the data from the first database and the data from the second database using the third graphical model.

Generating an expected prescriptions model using graphical models

Techniques disclosed herein include systems and methods for privacy-sensitive training data collection for updating acoustic models of speech recognition systems. In one embodiment, the system locally creates adaptation data from raw audio data. Such adaptation can include derived statistics and/or acoustic model update parameters. The derived statistics and/or updated acoustic model data can then be sent to a speech recognition server or third-party entity. Since the audio data and transcriptions are already processed, the statistics or acoustic model data is devoid of any information that could be human-readable or machine readable such as to enable reconstruction of audio data. Thus, such converted data sent to a server does not include personal or confidential information. Third-party servers can then continually update speech models without storing personal and confidential utterances of users.

Privacy-sensitive speech model creation via aggregation of multiple user models

Techniques disclosed herein include using a Maximum A Posteriori (MAP) adaptation process that imposes sparseness constraints to generate acoustic parameter adaptation data for specific users based on a relatively small set of training data. The resulting acoustic parameter adaptation data identifies changes for a relatively small fraction of acoustic parameters from a baseline acoustic speech model instead of changes to all acoustic parameters. This results in user-specific acoustic parameter adaptation data that is several orders of magnitude smaller than storage amounts otherwise required for a complete acoustic model. This provides customized acoustic speech models that increase recognition accuracy at a fraction of expected data storage requirements.

Sparse maximum a posteriori (map) adaption

Sparse maximum a posteriori (MAP) adaptation

Access is obtained to a large reference acoustic model for automatic speech recognition. The large reference acoustic model has L states modeled by L mixture models, and the large reference acoustic model has N components. A desired number of components N, less than N, to be used in a restructured acoustic model derived from the reference acoustic model, is identified. The desired number of components Nis selected based on a computing environment in which the restructured acoustic model is to be deployed. The restructured acoustic model also has L states. For each given one of the L mixture models in the reference acoustic model, a merge sequence is built which records, for a given cost function, sequential mergers of pairs of the components associated with the given one of the mixture models. A portion of the Ncomponents is assigned to each of the L states in the restructured acoustic model. The restructured acoustic model is built by, for each given one of the L states in the restructured acoustic model, applying the merge sequence to a corresponding one of the L mixture models in the reference acoustic model until the portion of the Ncomponents assigned to the given one of the L states is achieved.

Model restructuring for client and server based automatic speech recognition

Methods for compressing a transform associated with a feature space are presented. For example, a method for compressing a transform associated with a feature space includes obtaining the transform including a plurality of transform parameters, assigning each of a plurality of quantization levels for the plurality of transform parameters to one of a plurality of quantization values, and assigning each of the plurality of transform parameters to one of the plurality of quantization values to which one of the plurality of quantization levels is assigned. One or more of obtaining the transform, assigning of each of the plurality of quantization levels, and assigning of each of the transform parameters are implemented as instruction code executed on a processor device. Further, a Viterbi algorithm may be employed for use in non-uniform level/value assignments.

Compressing feature space transforms

A computer implemented method, system and/or computer program product confirm an orally entered address to a mobile navigation device. The mobile navigation device receives a global positioning system (GPS) root address component from a GPS. The GPS root address component is a text name of a root location at which a mobile navigation device is currently located. The mobile navigation device receives an orally entered address that comprises an oral root address component and an oral subunit component of the oral root address component. In response to the converted root address component matching the GPS root address component, the orally entered address is partitioned into the oral subunit component and the oral root address component, and any additional speech-to-text conversion of the orally entered address after the oral root address component is terminated.

Location based full address entry via speech recognition

A method, system, and computer program for class detection and time mediated averaging of class dependent models. A technique is described to take advantage of gender information in training data and how obtain female, male, and gender independent models from this information. By using a probability value to average male and female Gaussian Mixture Models (GMMs), dramatic deterioration in cross gender decoding performance is avoided.

Class detection scheme and time mediated averaging of class dependent models

A method, and a system to execute this method is being presented for the identification and separation of sources of an acoustic signal, which signal contains a mixture of multiple simultaneous component signals. The method represents the signal with multiple discrete state-variable sequences and combines acoustic and context level dynamics to achieve the source separation. The method identifies sources by discovering those frames of the signal whose features are dominated by single sources. The signal may be the simultaneous speech of multiple speakers.

System and method for speech separation and multi-talker speech recognition

In an Automatic Speech Recognition (ASR) system having at least two language models, a method is provided for combining language model scores generated by at least two language models. A list of most likely words is generated for a current word in a word sequence uttered by a speaker, and acoustic scores corresponding to the most likely words are also generated. Language model scores are computed for each of the most likely words in the list, for each of the at least two language models. A set of coefficients to be used to combine the language model scores of each of the most likely words in the list is respectively and dynamically determined, based on a context of the current word. The language model scores of each of the most likely words in the list are respectively combined to obtain a composite score for each of the most likely words in the list, using the set of coefficients determined therefor.

Dynamic language model mixtures with history-based buckets

Two statistics are disclosed for determining the quality of language models. These statistics are called acoustic perplexity and the synthetic acoustic word error rate (SAWER), and they depend upon methods for computing the acoustic confusability of words. It is possible to substitute models of acoustic data in place of real acoustic data in order to determine acoustic confusability. An evaluation model is created, a synthesizer model is created, and a matrix is determined from the evaluation and synthesizer models. Each of the evaluation and synthesizer models is a hidden Markov model. Once the matrix is determined, a confusability calculation may be performed. Different methods are used to determine synthetic likelihoods. The confusability may be normalized and smoothed and methods are disclosed that increase the speed of performing the matrix inversion and the confusability calculation. A method for caching and reusing computations for similar words is disclosed. Acoustic perplexity and SAWER are determined and applied.

Determining and using acoustic confusability, acoustic perplexity and synthetic acoustic word error rate

A characteristic-specific digitization method and apparatus are disclosed that reduces the error rate in converting input information into a computer-readable format. The input information is analyzed and subsets of the input information are classified according to whether the input information exhibits a specific physical parameter affecting recognition accuracy. If the input information exhibits the specific physical parameter affecting recognition accuracy, the characteristic-specific digitization system recognizes the input information using a characteristic-specific recognizer that demonstrates improved performance for the given physical parameter. If the input information does not exhibit the specific physical parameter affecting recognition accuracy, the characteristic-specific digitization system recognizes the input information using a general recognizer that performs well for typical input information. In one implementation, input speech having very low recognition accuracy as a result of a physical speech characteristic is automatically identified and recognized using a characteristic-specific speech recognizer.

Speech and signal digitization by using recognition metrics to select from multiple techniques

A parametric family of multivariate density functions formed by mixture models from univariate functions of the type exp(&minus;&verbar;x&verbar; ) for modeling acoustic feature vectores are used in automatic recognition of speech. The parameter &bgr; is used to measure the non-Gaussian nature of the data. &bgr; is estimated from the input data using a maximum likelihood criterion. There is a balance between &bgr; and the number of data points that must be satisfied for efficient estimation.

Impulsivity estimates of mixtures of the power exponential distrubutions in speech modeling

Growing community of inventors

New York, NY, United States of America

Peder Andreas Olsen