System and method for speech recognition modeling for mobile voice search

Patent metadata

Field	Value
Publication number	US-9558738-B2
Application number	US-201113042671-A
Country	US
Kind code	B2
Filing date	Mar 8, 2011
Priority date	Mar 8, 2011
Publication date	Jan 31, 2017
Grant date	Jan 31, 2017

Abstract

Disclosed herein are systems, methods, and non-transitory computer-readable storage media for generating an acoustic model for use in speech recognition. A system configured to practice the method first receives training data and identifies non-contextual lexical-level features in the training data. Then the system infers sentence-level features from the training data and generates a set of decision trees by node-splitting based on the non-contextual lexical-level features and the sentence-level features. The system decorrelates training vectors, based on the training data, for each decision tree in the set of decision trees to approximate full-covariance Gaussian models, and then can train an acoustic model for use in speech recognition based on the training data, the set of decision trees, and the training vectors.

First claim

We claim:

1. A method comprising:
receiving training data;
identifying non-contextual lexical-level features in the training data;
inferring sentence-level features from the training data using a sentence classifier seeded with portions of the training data;
estimating a first language model based on a sentence length by linearly interpolating subsets of the training data with a 3-gram language model, wherein the subsets are determined by length of sentences in the training data;
generating a set of decision trees by node-splitting according to the non-contextual lexical-level features and the sentence-level features;
decorrelating training vectors, according to the training data, for each decision tree in the set of decision trees to approximate full-covariance Gaussian models;
adapting a pre-existing acoustic model for use in speech recognition on received speech according to the training data, the set of decision trees, and the training vectors, to yield an adapted acoustic model;
combining the first language model with a second language model using interpolation weights and a frequency of carrier phrases in the training data, to yield a combined language model; and
performing the speech recognition on the received speech using the adapted acoustic model and the combined language model.

2. The method of claim 1 , wherein approximating the full-covariance Gaussian models is according to diagonal covariance.

3. The method of claim 1 , wherein the training data is unlabeled.

4. The method of claim 1 , wherein the node-splitting is further according to binary questions on the phonetic context of the training data.

5. The method of claim 4 , wherein the phonetic context is characterized by place and manner of articulation.

6. The method of claim 1 , wherein the pre-existing acoustic model is one of a triphone acoustic model and a pentaphone acoustic model.

7. The method of claim 1 , wherein the sentence-level features further comprise a gender and an accent of a speaker of the training data.

8. The method of claim 1 , wherein the non-contextual lexical-level features comprise one of a word-level feature and a syllable-level feature.

9. The method of claim 1 , wherein a sentence classifier infers sentence-level features from the training data.

10. The method of claim 9 , wherein the sentence classifier infers sentence-level features via a two-pass approach.

11. A system comprising:
a processor; and
a computer-readable storage medium having instructions stored which, when executed by the processor, cause the processor performing operations comprising:
receiving training data;
identifying non-contextual lexical-level features in the training data;
inferring sentence-level features from the training data using a sentence classifier seeded with portions of the training data;
estimating a first language model based on a sentence length by linearly interpolating subsets of the training data with a 3-gram language model, wherein the subsets are determined by length of sentences in the training data;
generating a set of decision trees by node-splitting according to the non-contextual lexical-level features and the sentence-level features;
decorrelating training vectors, according to the training data, for each decision tree in the set of decision trees to approximate full-covariance Gaussian models;
adapting a pre-existing acoustic model for use in speech recognition on received speech according to the training data, the set of decision trees, and the training vectors, to yield an adapted acoustic model;
combining the first language model with a second language model using interpolation weights and a frequency of carrier phrases in the training data, to yield a combined language model; and
performing the speech recognition on the received speech using the adapted acoustic model and the combined language model.

12. The system of claim 11 , wherein approximating the full-covariance Gaussian models is according to diagonal covariance.

13. The system of claim 11 , wherein the training data is unlabeled.

14. The system of claim 11 , wherein the node-splitting is further according to binary questions on the phonetic context of the training data.

15. The system of claim 14 , wherein the phonetic context is characterized by place and manner of articulation.

16. The system of claim 11 , wherein the pre-existing acoustic model is one of a triphone acoustic model and a pentaphone acoustic model.

17. A computer-readable storage device having instructions stored which, when executed by a computing device, cause the computing device to perform operations comprising:
receiving training data;
identifying non-contextual lexical-level features in the training data;
inferring sentence-level features from the training data using a sentence classifier seeded with portions of the training data;
estimating a first language model based on a sentence length by linearly interpolating subsets of the training data with a 3-gram language model, wherein the subsets are determined by length of sentences in the training data;
generating a set of decision trees by node-splitting according to the non-contextual lexical-level features and the sentence-level features;
decorrelating training vectors, according to the training data, for each decision tree in the set of decision trees to approximate full-covariance Gaussian models;
adapting a pre-existing acoustic model for use in speech recognition on received speech according to the training data, the set of decision trees, and the training vectors, to yield an adapted acoustic model;
combining the first language model with a second language model using interpolation weights and a frequency of carrier phrases in the training data, to yield a combined language model; and
performing the speech recognition on the received speech using the adapted acoustic model and the combined language model.

18. The computer-readable storage device of claim 17 , wherein the sentence-level features further comprise a gender and an accent of a speaker of the training data.

19. The computer-readable storage device of claim 17 , wherein the non-contextual lexical-level features comprise one of a word-level feature and a syllable-level feature.

Assignees

• Bocchieri Enrico
• Caseiro Diamantino Antonio
• Dimitriadis Dimitrios
• At & T Ip I Lp

Inventors

• Bocchieri Enrico
• Caseiro Diamantino Antonio
• Dimitriadis Dimitrios

Classifications

• G10L15/063 Primary

  Training · CPC title

• G10L15/14

  using statistical models, e.g. Hidden Markov Models [HMMs] (G10L15/18 takes precedence) · CPC title