Matsushita Electric Industrial Co., Ltd.

Panasonic Technologies, Inc.

Yunxin Zhao

Hisashi Wakita

Yoshio Ono

A speaker adaptation technique based on the separation of speech spectra variation sources is developed for improving speaker-independent continuous speech recognition. The variation sources include speaker acoustic characteristics, and contextual dependency of allophones. Statistical methods are formulated to normalize speech spectra based on speaker acoustic characteristics and then adapt mixture Gaussian density phone models based on speaker phonologic characteristics. Adaptation experiments using short calibration speech (5 sec./speaker) have shown substantial performance improvement over the baseline recognition system.

Self-learning speaker adaptation based on spectral bias source

A speaker normalization method is described based on spectral shifts in the auditory filter domain. The method is characterized by using an estimated vocal tract length as a criterion to determine the spectral shift value. Certain constraints are found to be necessary for the shift in the auditory filter domain, and two techniques based on these constraints, the One-Bark shift and the refined Bark-scale shift, are introduced. When tested in vowel classification experiments, significant performance improvement was obtained for both techniques. The method is useful for speaker normalization in speaker-independent speech recognition.

Speaker normalization using constrained spectra shifts in auditory

A self-learning speaker adaptation method for automatic speech recognition is provided. The method includes building a plurality of Gaussian mixture density phone models for use in recognizing speech. The Gaussian mixture density phone models are used to recognize a first utterance of speech from a given speaker. After the first utterance of speech has been recognized, the recognized first utterance of speech is used to adapt the Gaussian mixture density hone models for use in recognizing a subsequent utterance of speech from that same speaker, whereby the Gaussian mixture density phone models are automatically adapted to that speaker in self-learning fashion to thereby produce a plurality of adapted Gaussian mixture density phone models.

Self-learning speaker adaptation based on spectral variation source

A model-training module generates mixture Gaussian density models from speech training data for continuous, or isolated word speech recognition systems. Speech feature sequences are labeled into segments of states of speech units using Viterbi-decoding based optimized segmentation algorithm. Each segment is modeled by a Gaussian density, and the parameters are estimated by sample mean and sample covariance. A mixture Gaussian density is generated for each state of each speech unit by merging the Gaussian densities of all the segments with the same corresponding label. The resulting number of mixture components is proportional to the dispersion and sample size of the training data. A single, fully merged, Gaussian density is also generated for each state of each speech unit. The covariance matrices of the mixture components are selectively smoothed by a measure of relative sharpness of the Gaussian density and the smoothing can also be done blockwise. The weights of the mixture components are set uniformly initially, and are reestimated using a segmental-average procedure. The weighting coefficients, together with the Gaussian densities, then become the models of speech units for use in speech recognition.

Training module for estimating mixture Gaussian densities for speech

A word hypothesis module for speech decoding consists of four submodules: vowel center detection, bidirectional tree searches around each vowel center, forward-backward pruning, and additional short words hypotheses. By detecting the strong energy vowel centers, a vowel-centered lexicon tree can be placed at each vowel center and searches can be performed in both the left and right directions, where only simple phone models are used for fast acoustic match. A stage-wise forward-backward technique computes the word-beginning and word-ending likelihood scores over the generated half-word lattice for further pruning of the lattice. To avoid potential miss of short words with weak energy vowel centers, a lexicon tree is compiled for these words and tree searches are performed between each pair of adjacent vowel centers. The integration of the word hypothesizer with a top-down Viterbi beam search in continuous speech decoding provides two-pass decoding which significantly reduces computation time.

Word hypothesizer for continuous speech decoding using stressed-vowel

A model-training module generates mixture Gaussian density models from speech training data for continuous, or isolated word HMM-based speech recognition systems. Speech feature sequences are labeled into segments of states of speech units using Viterbi-decoding based optimized segmentation algorithm. Each segment is modeled by a Gaussian density, and the parameters are estimated by sample mean and sample covariance. A mixture Gaussian density is generated for each state of each speech unit by merging the Gaussian densities of all the segments with the same corresponding label. The resulting number of mixture components is proportional to the dispension and sample size of the training data. A single, fully merged, Gaussian density is also generated for each state of each speech unit. The covariance matrices of the mixture components are selectively smoothed by a measure of relative sharpness of the Gaussian density. The weights of the mixture components are set uniformly initially, and are reestimated using a segmental-average procedure. The weighting coefficients, together with the Gaussian densities, then become the models of speech units for use in speech recognition.

Growing community of inventors

Santa Barbara, CA, United States of America

Yunxin Zhao