Qualcomm Incorporated

U S West Advanced Technologies, Inc.

International Computer Science Institute

The Johns Hopkins University

Columbia University

Hynek Hermansky

Harinath Garudadri

Petr Motlicek

Nelson H Morgan

Kenneth W Church

Keith Kintzley

Aren Jansen

Naveen Srinivasamurthy

Sangita R Sharma

Louis A Cox, Jr

Daniel Ellis

Daniel P W Ellis

Naveen B Srinivasamurthy

Philip D Kohn

Marios Athineos

Maria Carmen Benitez Ortuzar

Sriram Ganapathy

Lukas Burget

Stephane N Dupont

Frantisek Grezl

Pratibha Jain

Sachin Kajarekar

Sunil Sivadas

Charles C Wooters

Andre Gustavo Adami

A system and method are provided for performing speech processing. A system includes an audio detection system configured to receive a signal including speech and a memory having stored therein a database of keyword models forming an ensemble of filters associated with each keyword in the database. A processor is configured to receive the signal including speech from the audio detection system, decompose the signal including speech into a sparse set of phonetic impulses, and access the database of keywords and convolve the sparse set of phonetic impulses with the ensemble of filters. The processor is further configured to identify keywords within the signal including speech based a result of the convolution and control operation the electronic system based on the keywords identified.

System and method for processing speech to identify keywords or other information

A technique of spectral noise shaping in an audio coding system is disclosed. Frequency decomposition of an input audio signal is performed to obtain multiple frequency sub-bands that closely follow critical bands of human auditory system decomposition. The tonality of each sub-band is determined. If a sub-band is tonal, time domain linear prediction (TDLP) processing is applied to the sub-band, yielding a residual signal and linear predictive coding (LPC) coefficients of an all-pole model representing the sub-band signal. The residual signal is further processed using a frequency domain linear prediction (FDLP) method. The FDLP parameters and LPC coefficients are transferred to a decoder. At the decoder, an inverse-FDLP process is applied to the encoded residual signal followed by an inverse TDLP process, which shapes the quantization noise according to the power spectral density of the original sub-band signal. Non-tonal sub-band signals bypass the TDLP process.

Spectral noise shaping in audio coding based on spectral dynamics in frequency sub-bands

In an apparatus and method, time-varying signals are processed and encoded via a frequency domain linear prediction (FDLP) scheme to arrive at an all-pole model. Residual signals resulted from the scheme are estimated and transformed into a time domain signal. Through the process of heterodyning, the time domain signal is frequency shifted toward the baseband level as a downshifted carrier signal. Quantized values of the all-pole model and the frequency transform of the downshifted carrier signal are packetized as encoded signals suitable for transmission or storage. To reconstruct the time-varying signals, the encoded signals are decoded. The decoding process is basically the reverse of the encoding process.

Processing of excitation in audio coding and decoding

In an apparatus and method, time-varying signals are processed and encoded via a frequency domain linear prediction (FDLP) scheme to arrive at an all-pole model. Residual signals resulted from the scheme are estimated. Quantized values of the all-pole model and the residual signals are packetized as encoded signals suitable for transmission or storage. To reconstruct the time-varying signals, the encoded signals are decoded. The decoding process is basically the reverse of the encoding process.

Signal coding and decoding based on spectral dynamics

In accordance with the present invention, computer implemented methods and systems are provided for representing and modeling the temporal structure of audio signals. In response to receiving a signal, a time-to-frequency domain transformation on at least a portion of the received signal to generate a frequency domain representation is performed. The time-to-frequency domain transformation converts the signal from a time domain representation to the frequency domain representation. A frequency domain linear prediction (FDLP) is performed on the frequency domain representation to estimate a temporal envelope of the frequency domain representation. Based on the temporal envelope, one or more speech features are generated.

Systems and methods for speech recognition using frequency domain linear prediction polynomials to form temporal and spectral envelopes from frequency domain representations of signals

The present invention successfully combines neural-net discriminative feature processing with Gaussian-mixture distribution modeling (GMM). By training one or more neural networks to generate subword probability posteriors, then using transformations of these estimates as the base features for a conventionally-trained Gaussian-mixture based system, substantial error rate reductions may be achieved. The present invention effectively has two acoustic models in tandem&#x2014;first a neural net and then a GMM. By using a variety of combination schemes available for connectionist models, various systems based upon multiple features streams can be constructed with even greater error rate reductions.

Nonlinear mapping for feature extraction in automatic speech recognition

A distributed voice recognition system and method for obtaining acoustic features and speech activity at multiple frequencies by extracting high frequency components thereof on a device, such as a subscriber station and transmitting them to a network server having multiple stream processing capability, including cepstral feature processing, MLP nonlinear transformation processing, and multiband temporal pattern architecture processing. The features received at the network server are processed using all three streams, wherein each of the three streams provide benefits not available in the other two, thereby enhancing feature interpretation. Feature extraction and feature interpretation may operate at multiple frequencies, including but not limited to 8 kHz, 11 kHz, and 16 kHz.

Multistream network feature processing for a distributed speech recognition system

A method and system are provided for alleviating the harmful effects of convolutional and additive noise in speech, such as due to the environmental noise and linear spectral modification, based on the filtering of time trajectories of an auditory-like spectrum in a particular spectral domain.

Noise resistant auditory model for parametrization of speech

A method and system are provided for alleviating the harmful effects of convolutional distortions of speech, such as the effect of a telecommunication channel, on the performance of an automatic speech recognizer (ASR). The technique is based on the filtering of time trajectories of an auditory-like spectrum derived from the Perceptual Linear Predictive (PLP) method of speech parameter estimation.

Auditory model for parametrization of speech

A method for synthesizing human speech using a linear mapping of a small set of coefficients that are speaker-independent. Preferably, the speaker-independent set of coefficients are cepstral coefficients developed during a training session using a perceptual linear predictive analysis. A linear predictive all-pole model is used to develop corresponding formants and bandwidths to which the cepstral coefficients are mapped by using a separate multiple regression model for each of the five formant frequencies and five formant bandwidths. The dual analysis produces both the cepstral coefficients of the PLP model for the different vowel-like sounds and their true formant frequencies and bandwidths. The separate multiple regression models developed by mapping the cepstral coefficients into the formant frequencies and formant bandwidths can then be applied to cepstral coefficients determined for subsequent speech to produce corresponding formants and bandwidths used to synthesize that speech. Since less data are required for synthesizing each speech segment than in conventional techniques, a reduction in the required storage space and/or transmission rate for the data required in the speech synthesis is achieved. In addition, the cepstral coefficients for each speech segment can be used with the regressive model for a different speaker, to produce synthesized speech corresponding to the different speaker.

Speech synthesis using perceptual linear prediction parameters

Growing community of inventors

Baltimore, MD, United States of America

Hynek Hermansky