Texas Instruments Corporation

Yifan Gong

Alexis P Bernard

Yu-Hung Kao

Coimbatore S Ramalingam

John J Godfrey

A method for performing time and frequency Signal-to-Noise Ratio (SNR) dependent weighting in speech recognition is described that includes for each period t estimating the SNR to get time and frequency SNR information &#x3b7;; calculating the time and frequency weighting to get &#x3b3;; performing the back and forth weighted time varying DCT transformation matrix computation MGMto get T; providing the transformation matrix Tand the original MFCC feature othat contains the information about the SNR to a recognizer including the Viterbi decoding; and performing weighted Viterbi recognition b(o).

Middle-end solution to robust speech recognition

For a given sentence grammar, speech recognizers are often required to decode M sets of HMMs each of which models a specific acoustic environment. In order to match input acoustic observations to each of the environments, typically recognition search methods require a network of M sub-networks. A new speech recognition search method is described here, which needs a network that is only the size of a single sub-network and yet provides the same recognition performance, thus reducing the memory requirements for network storage by (M-1)/M.

Decoding multiple HMM sets using a single sentence grammar

A method and detector for providing a noise resistant utterance detector is provided by extracting a noise estimate () to augment the signal-to-noise ratio of the speech signal, inverse filtering () of the speech signal to focus on the periodic excitation part of the signal and spectral reshaping () to accentuate separation between formants.

Noise-resistant utterance detector

For a given sentence grammar, speech recognizers are often required to decode M sets of HMMs each of which models a specific acoustic environment. In order to match input acoustic observations to each of the environments, typically recognition search methods require a network of M sub-networks. A new speech recognition search method is described here, which needs that is only the size of a single sub-network and yet gives the same recognition performance, thus reducing memory requirement for network storage by (M&#x2212;1)/M.

The operating range of joint additive and convolutive compensating method is extended by enhanced channel estimation procedure that adds SNR-dependent inertia and SNR-dependent limit on the channel estimate.

Method to extend operating range of joint additive and convolutive compensating algorithms

A speech recognizer operating in both ambient noise (additive distortion) and microphone changes (convolutive distortion) is provided. For each utterance to be recognized the recognizer system adapts HMM mean vectors with noise estimates calculated from pre-utterance pause and a channel estimate calculated using an Estimation Maximization algorithm from previous utterances.

Method of speech recognition resistant to convolutive distortion and additive distortion

A small vocabulary speech recognizer suitable for implementation on a 16-bit fixed-point DSP is described. The input speech xt is sampled at analog-to-digital (A/D) converterand the digital samples are applied to MFCC (Mel-scaled cepstrum coefficients) front end processing. For robustness to background noises, PMC (parallel model combination)is integrated. The MFCC and Gaussian mean vectors are applied to PMC. The MFCC and PMC provide speech features extracted in noise and this is used to modify the HMMs. The noise adapted HMMs excluding mean vectors are applied to the search procedure to recognize the grammar. A method of computing MFCC comprises the steps of: performing dynamic Q-point computation for the preemphasis, Hamming Window, FFT, complex FFT to power spectrum and Mel scale power spectrum into filter bank steps, a log filter bank step and after the log filter bank step performing fixed Q-point computation. A polynomial fit is used to compute log2 in the log filter bank step. The method of computing PMC comprises the steps of: computing noise MFCC profile, computing cosine transform MFCC into mel-scale filter bank, converting log filter bank into linear filter bank with an exponential wherein to compute exp2 a polynomial fit is used, performing a model combination in the linear filter bank domain; and converting the noise compensated linear filter bank into MFCC by log and inverse cosine transform.

Implementing a high accuracy continuous speech recognizer on a fixed-point processor

A new iterative hierarchical linear regression method for generating a set of linear transforms to adapt HMM speech models to a new environment for improved speech recognition is disclosed. The method determines a new set of linear transforms at an iterative step by Estimate-Maximize (EM) estimation, and then combines the new set of linear transforms with the prior set of linear transforms to form a new merged set of linear transforms. An iterative step may include realignment of adaptation speech data to the adapted HMM models to further improve speech recognition performance.

Accumulating transformations for hierarchical linear regression HMM adaptation

A method of speech recognition with compensation is provided by modifying HMM models trained on clean speech with cepstral mean normalization. For all speech utterances the MFCC vector is calculated for the clean database. This mean MFCC vector is added to the original models. An estimate of the background noise is determined for a given speech utterance. The model mean vectors adapted to the noise are determined. The mean vector of the noisy data over the noisy speech space is determined and this is removed from model mean vectors adapted to noise to get the target model.

Method of speech recognition with compensation for both channel distortion and background noise

An utterance detector for speech recognition is described. The detector consists of two components. The first part makes a speech/non-speech decision for each incoming speech frame. The decision is based on a frequency-selective autocorrelation function obtained by speech power spectrum estimation, frequency filter, and inverse Fourier transform. The second component makes utterance detection decision, using a state machine that describes the detection process in terms of the speech/non-speech decision made by the first component.

Automatic utterance detector with high noise immunity

A maximum likelihood (ML) linear regression (LR) solution to environment normalization is provided where the environment is modeled as a hidden (non-observable) variable. By application of an expectation maximization algorithm and extension of Baum-Welch forward and backward variables (Steps) a source normalization is achieved such that it is not necessary to label a database in terms of environment such as speaker identity, channel, microphone and noise type.

Source normalization training for HMM modeling of speech

A method and system for calibration of a data acquisition path is achieved by applying a voice utterance to a first high quality microphone and reference path and to a test acquisition path including a test microphone such as a lower quality one used in a car. The calibration device includes detecting the power density of the reference signal Ythrough the reference path and detecting the power density of the signal Ythrough the acquisition path. A processor processes these signals to provide an output signal representing a noise estimate and channel estimate. The processing uses equation derived by modeling convolutive and additive noise as polynomials with different orders and estimating model parameters using maximum likelihood criterion and simultaneously solving linear equations for the different orders.

Calibration of speech data acquisition path

On improved transformation method uses an initial set of Hidden Markov Models (HMMs) trained on a large amount of speech recorded in a low noise environment R to provide rich information on co-articulation and speaker variation and a smaller database in a more noisy target environment T. A set H of HMMs is trained with data provided in the low noise environment R and the utterances in the noisy environment T are transcribed phonetically using set H of HMMs. The transcribed segments are grouped into a set of Classes C. For each subclass c of Classes C, the transformation &PHgr; is found to maximize likelihood utterances in T, given H. The HMMs are transformed and steps repeated until likelihood stabilizes.

Method for transforming HMMs for speaker-independent recognition in a noisy environment

Reducing mismatch between HMMs trained with clean speech and speech signals recorded under background noise can be approached by distribution adaptation using parallel model combination (PMC). Accurate PMC has no closed-form expression, therefore simplification assumptions must be made in implementation. Under a new log-max assumption, adaptation formula for log-spectral parameters are presented, both for static and dynamic parameters. The system takes the mean vector ( ) made up of the static part ( ) and the dynamic part ( ) and the noise vector ( ) made up of the static part ( ) and dynamic part ( ) and applies to a decision circuit ( ) to determine if the quiet vector plus the gain of speech produced in noise with respect to clean speech is greater than the noisy mean vector and if so the static part is equal to the gain plus the clean speech vector and the dynamic part is the change in the the quiet speech vector and if not greater than the noise vector then the static part equals the noise vector and the dynamic part is zero.

Log-spectral compensation of PMC Gaussian mean vectors for noisy speech recognition using log-max assumption

An estimate of clean speech vector, typically Mel-Frequency Cepstral Coefficient (MFCC) given its noisy observation is provided. The method makes use of two Gaussian mixtures. The first one is trained on clean speech and the second is derived from the first one using some noise samples. The method gives an estimate of a clean speech feature vector as the conditional expectancy of clean speech given an observed noisy vector.

Growing community of inventors

Plano, TX, United States of America

Yifan Gong