System and method for learning alternate pronunciations for speech recognition

Patent metadata

Field	Value
Publication number	US-9489943-B2
Application number	US-201414515607-A
Country	US
Kind code	B2
Filing date	Oct 16, 2014
Priority date	Oct 16, 2013
Publication date	Nov 8, 2016
Grant date	Nov 8, 2016

Abstract

A system and method for learning alternate pronunciations for speech recognition is disclosed. Alternative name pronunciations may be covered, through pronunciation learning, that have not been previously covered in a general pronunciation dictionary. In an embodiment, the detection of phone-level and syllable-level mispronunciations in words and sentences may be based on acoustic models trained by Hidden Markov Models. Mispronunciations may be detected by comparing the likelihood of the potential state of the targeting pronunciation unit with a pre-determined threshold through a series of tests. It is also within the scope of an embodiment to detect accents.

First claim

The invention claimed is:

1. A method for learning pronunciation in a given language comprising the steps of:
a. training an acoustic model on a large speech corpus to distinguish phonemes;
b. constructing a phoneme confusion matrix;
c. constructing a phoneme replacement candidate list for each phoneme in a set of speech data containing pronunciations for recognition;
d. learning alternative pronunciations of a word that has been mispronounced;
e. combining said learned alternative pronunciations with a linguistic dictionary to create a pooled dictionary; and
f. pruning said pooled dictionary to limit the number of learned alternative pronunciations in order to create an improved dictionary.

2. The method of claim 1 , wherein the acoustic model in step (a) is trained by one of: maximum likelihood criterion and discriminative training criterion.

3. The acoustic model of claim 1 , wherein said acoustic model in step (a) is based on a Hidden Markov Model and Gaussian Mixture Model.

4. The method of claim 1 , wherein step (b) further comprises the step of merging an acoustic confusion matrix with a linguistic confusion matrix to construct said phoneme confusion matrix.

5. The method of claim 4 , wherein said acoustic confusion matrix is obtained by performing a phoneme recognition experiment on a test set of speech data.

6. The acoustic confusion matrix of claim 5 , wherein a low value in the acoustic confusion matrix indicates a phoneme is similar to other phonemes and confusable.

7. The linguistic confusion matrix of claim 4 , wherein the linguistic confusion matrix is provided by a group of linguistic experts.

8. The linguistic confusion matrix of claim 7 , wherein the linguistic confusion matrix comprises a binary matrix comprising the numbers 0 and 1.

9. The linguistic confusion matrix of claim 8 , wherein 0 indicates that a phoneme belongs to the same confusion cluster as an other phoneme and the phonemes are confusable.

10. The method of claim 1 , wherein step (c) further comprises the steps of:
a. selecting a phoneme from the speech data as a target phoneme for analysis and arranging the remaining phonemes based on distance to the target phoneme;
b. applying a statistical clustering algorithm to similarly group the arranged phonemes;
c. constructing the list of phoneme replacement candidates for the target phoneme from the similarly grouped phonemes; and
d. repeating all of the steps for each phoneme in the speech data set.

11. The method of claim 10 , wherein the distance between a phoneme and said target phoneme in step (a) represents a confusion value in a phoneme confusion matrix.

12. The phoneme confusion matrix of claim 11 , wherein a low value indicates high confusion between a phoneme and a target phoneme.

13. The method of claim 1 , wherein step (d) further comprises the steps of:
a. obtaining the original pronunciation for each word that has been misrecognized;
b. generating an alternative pronunciation data set wherein an improved new pronunciation is compared with an original pronunciation;
c. performing recognition on the alternative pronunciation data set, the acoustic model, and the set of speech data;
d. determining the best pronunciation from the alternative pronunciation data set;
e. retaining selected pronunciations from said alternative pronunciation data set; and
f. repeating the steps for all misrecognized words to form the learned pronunciation data set.

14. The method of claim 13 , wherein the original pronunciation in step (a) is obtained from one of: a linguistic dictionary and an automatic word-to-phoneme generator.

15. The method of claim 13 , wherein step (b) further comprises the steps of:
a. placing groups of phonemes in their respective positions; and
b. obtaining all of the phoneme combinations.

16. The determination of the best pronunciation of claim 13 , wherein the best pronunciation of the word in step (d) results in the highest recognition accuracy.

17. The method of claim 16 , wherein step (b) further comprises the step of: determining the size of the alternative pronunciation data set by the mathematical equation:
X=Π m=1 M N m .

18. The method of claim 13 , wherein the recognition of step (c) is performed using a Viterbi decoding algorithm.

19. The linguistic dictionary of claim 1 , wherein the linguistic dictionary comprises a set of pronunciations of common words in a language and is provided by a group of linguistic experts.

20. The method of claim 1 , wherein creation of the improved dictionary in step (f) further comprises the steps of:
a. computing a distance from each word to an other word in the linguistic dictionary;
b. creating a subset of similar words for each misrecognized word;
c. performing recognition on the subset of similar words, the acoustic model and the set of speech;
d. identifying the frequency of failure;
e. removing a pronunciation contributing to frequency failure greater than a threshold; and
f. repeating the process for all misrecognized words.

21. The method of claim 20 , wherein computing of step (a) comprises a dynamic programming algorithm based on word pronunciations.

22. The method of claim 21 , wherein the dynamic programming algorithm uses the accumulative confusion values in the phoneme confusion matrix as costs in computation.

23. The confusion values of claim 22 , wherein a low value indicates word similarity.

24. The method of claim 20 , wherein step (b) further comprises the step of selecting a number of words based on a threshold.

25. The method of claim 20 , wherein step (d) further comprises the step of identifying incorrect recognitions, a pronunciation associated with an incorrect recognition, and the frequency of failure related to an incorrect recognition.

26. The method of claim 20 , wherein the recognition of step (c) is performed using a Viterbi decoding algorithm.

27. The method of claim 1 , wherein step (f) further comprises the step of: applying the improved dictionary in a grammar based speech recognition task to improve speech recognition accuracy.

28. The method of claim 1 further comprising the step of optimizing the efficiency of learning alternative pronunciations, wherein said optimizing comprises one or more of the following:
a. reducing the length of a phoneme replacement candidate list for each phoneme in the original pronunciation of the a word that has been mispronounced if a number of candidate pronunciations exceed a threshold; and
b. optimizing a phoneme determination order when obtaining a desired pronunciation for a misrecognized word.

29. The method of claim 28 , wherein step (a) further comprises the step of determining the scale of length reduction of a phoneme replacement candidate list with the mathematical equation:

r
m
′

=

(

M

ma
⁢

⁢
x

Assignees

• Interactive Intelligence Group Inc

Inventors

• Ge Zhenhao
• Tyagi Vivek
• Ganapathiraju Aravind
• Iyer Ananth Nagaraja
• Randal Scott Allen
• Wyss Felix Immanuel

Classifications

• G10L2015/081

  Search algorithms, e.g. Baum-Welch or Viterbi · CPC title

• G10L15/187 Primary

  Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams · CPC title

• G09B19/06

  Foreign languages (with audible presentation of material to be studied G09B5/04) · CPC title

• G10L15/063 Primary

  Training · CPC title

• G10L15/14

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