Speech recognition method and system

• US 7,319,960 B2
• Filed: 12/19/2001
• Issued: 01/15/2008
• Est. Priority Date: 12/19/2000
• Status: Expired due to Fees

• Current Assignee: Nokia Corporation
• Original Assignee: Nokia Corporation
• Inventors: Koumpis, Konstantinos, Riis, Soren
• Primary Examiner(s): Hudspeth; David
• Assistant Examiner(s): Ng; Eunice
• Application Number:
  US10/020,895
• Publication Number:
  US 20020165715A1
• Time in Patent Office: 2,218 Days
• Field of Search: 704/253, 704/254, 704/256, 704/256.1, 704/256.4, 704/255, 704/232, 704/251
• US Class Current: 704/256.4
• CPC Class Codes: G10L 15/08 Speech classification or search

Abstract

A speech recognition system uses a phoneme counter to determine the length of a word to be recognized. The result is used to split a lexicon into one or more sub-lexicons containing only words which have the same or similar length to that of the word to be recognized, so restricting the search space significantly. In another aspect, a phoneme counter is used to estimate the number of phonemes in a word so that a transition bias can be calculated. This bias is applied to the transition probabilities between phoneme models in an HNN based recognizer to improve recognition performance for relatively short or long words.

Claims

1. A speech recognition system in which a word to be recognized is represented as a sequence of phonetic segment models in which a transition probability represents the probability of the occurrence of a transition between the models, comprising:

means for estimating the number of phonetic segments in the word to be recognized; and

means for biasing the transition probabilities in dependence on the estimated number of phonetic segments in the word.

2. A speech recognition system according to claim 1, wherein the biasing means comprise means for applying a transition bias to each of the transition probabilities between a plurality of phonetic segment models.

3. A speech recognition system according to claim 2, operable to recognize words from a recognition vocabulary, wherein the transition bias is calculated as the transition bias which maximizes recognition performance on a validation data set which represents the recognition vocabulary.

4. A speech recognition system according to claim 3, wherein the validation data set has the same vocabulary as the recognition vocabulary.

5. A speech recognition system according to claim 1, wherein the estimating means comprises a speaker specific rate of speech estimator.

6. A speech recognition system according to claim 1, wherein the estimating means comprises a Free Order Viterbi decoder.

7. A speech recognition system according to claim 1, wherein the estimating means comprises a neural network classifier.

8. A speech recognition system according to claim 1, wherein the transition bias is set in response to the result of the estimating means.

9. A speech recognition system according to claim 8, comprising table look-up means for setting the transition bias in accordance with the number of phonetic segments in the word.

10. A speech recognition system according to claim 8, comprising direct setting means for setting the transition bias as proportional to the number of phonetic segments in the word.

11. A speech recognition system according to claim 10, wherein the direct setting means is arranged to set the transition bias to be equal to the number of phonetic segments in the word.

12. A speech recognition system according to claim 1, wherein each phonetic segment comprises a phoneme.

13. A speech recognition system according to claim 1, further comprising means for performing word recognition for the word on an individual basis based on the biased transition probabilities.

14. A speech recognition system according to claim 13, further comprising means for performing word recognition for each word in a multiword sentence based on a biased transition probability determined separately for each corresponding word in the sentence based on the estimated number of phonetic segments in the each corresponding word.

15. A speech recognition system in which a word to be recognized is represented as a sequence of phonetic segment models in which a transition probability represents the probability of the occurrence of a transition between the models, comprising:

a phonetic segment estimator arranged to output an estimate of the number of phonetic segments in the word; and

a processing module for applying a transition bias to the transition probability in dependence on the estimate of the number of phonetic segments in the word.

16. A portable communications device including a speech recognition system according to claim 15.

17. A speech recognition system according to claim 15, wherein the processing module is configured to perform word recognition for the word on an individual basis based on the biased transition probabilities.

18. A speech recognition system according to claim 17, wherein the processing module is configured to perform word recognition for each word in a multiword sentence based on a biased transition probability determined separately for each corresponding word in the sentence based on the estimated number of phonetic segments in the each corresponding word.

19. A method of speech recognition in which a word to be recognized is represented as a sequence of phonetic segment models in which a transition probability represents the probability of the occurrence of a transition between the models, the method comprising biasing the transition probabilities in dependence of the number of phonetic segments in the word.

20. A method according to claim 19, comprising decoding the sequence of phonetic segment models after application of the transition bias.

21. A method according to claim 19, comprising decoding the sequence of phonetic segment models without the application of transition bias and normalizing the resulting scores by a contribution proportional to the transition bias.

22. A method according to claim 21, comprising calculating the transition bias in parallel with the decoding of the sequence of phonetic segment models.

23. A method according to claim 19, further comprising performing word recognition for the word on an individual basis based on the biased transition probabilities.

24. A method according to claim 23, further comprising performing word recognition for each word in a multiword sentence based on a biased transition probability determined separately for each corresponding word in the sentence based on the estimated number of phonetic segments in the each corresponding word.

25. A method comprising:

receiving a word to be recognized represented as a sequence of phonetic segment models in which a transition probability represents the probability of the occurrence of a transition between the models;

biasing the transition probabilities in dependence of the number of phonetic segments in the word; and

performing word recognition for the word on an individual basis based on the biased transition probabilities.

26. A method according to claim 25, further comprising performing word recognition for each word in a multiword sentence based on a biased transition probability determined separately for each corresponding word in the sentence based on the estimated number of phonetic segments in the each corresponding word.

27. An apparatus comprising a processing element configured to:

estimate a number of phonetic segments in a word to be recognized; and

bias transition probabilities in dependence on the estimated number of phonetic segments in the word to be recognized,

wherein the word to be recognized is represented as a sequence of phonetic segment models in which a transition probability represents the probability of the occurrence of a transition between the models.