Utterance verification method and apparatus for isolated word N-best recognition result

• US 8,374,869 B2
• Filed: 08/04/2009
• Issued: 02/12/2013
• Est. Priority Date: 12/22/2008
• Status: Expired due to Fees

• Current Assignee: Electronics and Telecommunications Research Institute
• Original Assignee: Electronics and Telecommunications Research Institute
• Inventors: Kang, Jeom Ja, Lee, Yunkeun, Park, Jeon Gue, Jung, Ho-Young, Jeon, Hyung-Bae, Chung, Hoon, Lee, Sung Joo, Chung, Euisok, Wang, Ji Hyun, Kang, Byung Ok, Park, Ki-young, Kim, Jong Jin
• Primary Examiner(s): GUERRA-ERAZO, EDGAR X
• Application Number:
  US12/535,024
• Publication Number:
  US 20100161334A1
• Time in Patent Office: 1,288 Days
• Field of Search: 704/233, 704/243, 704/251, 704/255, 704/256, 704/8, 704/252, 704/253, 704/254, 704/256.4, 704/231, 704/235, 704/246, 704/250, 704/241, 704/245, 704/248
• US Class Current: 704/251
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

An utterance verification method for an isolated word N-best speech recognition result includes: calculating log likelihoods of a context-dependent phoneme and an anti-phoneme model based on an N-best speech recognition result for an input utterance; measuring a confidence score of an N-best speech-recognized word using the log likelihoods; calculating distance between phonemes for the N-best speech-recognized word; comparing the confidence score with a threshold and the distance with a predetermined mean of distances; and accepting the N-best speech-recognized word when the compared results for the confidence score and the distance correspond to acceptance.

Claims

1. An utterance verification method for an isolated word N-best speech recognition result, comprising:

calculating log likelihoods of a context-dependent phoneme and an anti-phoneme model based on an N-best speech recognition result for an input utterance;

measuring a confidence score of an N-best speech-recognized word using the log likelihoods;

calculating a distance between phonemes for the N-best speech-recognized word;

comparing the confidence score with a threshold and the distance with a mean of distances; and

accepting the N-best speech-recognized word when the compared results for the confidence score and the distance correspond to acceptance,

wherein the log likelihood of the context-dependent phoneme is calculated by;

log likelihood of context-dependent phoneme={(log likelihood of current phoneme)−

(mean of base phonemes of current phoneme)}/(standard deviation of base phonemes of current phoneme).

2. The utterance verification method of claim 1, wherein said comparing includes:

accepting the N-best speech-recognized word when the confidence score is greater than the threshold;

rejecting the N-best speech-recognized word when the confidence score is less than the threshold;

accepting the N-best speech-recognized word when the distance is less than the mean of distances; and

rejecting the N-best speech-recognized word when the distance is greater than the mean of distances.

3. The utterance verification method of claim 1, further comprising:

rejecting the N-best speech-recognized word when the compared results for the confidence score and the distance correspond to rejections.

4. The utterance verification method of claim 1, further comprising:

determining indetermination for the N-best speech-recognized word, when any one of the compared results for the confidence score and the distance corresponds to acceptance and the other corresponds to rejection.

5. The utterance verification method of claim 1,

wherein the confidence score of the N-best speech-recognized word is calculated by:

word-level confidence score={(log likelihood of the N-best speech recognized word for the context-dependent phoneme model)−

(log likelihood of the N-best speech recognized word for the anti-phoneme model word)}/(the number of phonemes in the N-best speech recognized word).

6. The utterance verification method of claim 5, wherein the mean of distances is calculated by:

mean of distances=(sum of distances of N-best recognition results)/the number of N-best.

7. The utterance verification method of claim 1, wherein the N-best speech recognition is performed by Viterbi search by referring to a context-dependent phoneme model extracted from a feature vector of the input utterance.

8. An utterance verification method for an isolated word N-best speech recognition result, comprising:

calculating log likelihoods of a context-dependent phoneme and an anti-phoneme model based on an N-best speech recognition result for an input utterance;

measuring a confidence score of an N-best speech-recognized word using the log likelihoods;

calculating a distance between phonemes for the N-best speech-recognized word;

comparing the confidence score with a threshold and the distance with a mean of distances; and

accepting the N-best speech-recognized word when the compared results for the confidence score and the distance correspond to acceptance,

wherein the log likelihood of the anti-phoneme model is calculated by;

log likelihood of anti-phoneme model={(log likelihood of anti-phoneme model of current phoneme)−

(mean of anti-phoneme model of current phoneme)}/(standard deviation of anti-phoneme model of current phoneme).

9. An utterance verification system for an isolated word N-best speech recognition result comprising:

a computer comprising;

a pre-processor extracting a feature vector of an input utterance and performing endpoint detection;

an N-best speech recognizer performing N-best speech recognition through Viterbi search by referring to the context-dependent phoneme model extracted from the feature vector; and

an N-best utterance verifier calculating log likelihoods of a context-dependent phoneme and an anti-phoneme model for the N-best speech-recognized word, comparing a confidence score measured for the N-best speech-recognized word with a threshold, comparing a distance measured for the N-best speech-recognized-word with a mean of distances, and accepting the N-best speech-recognized word when the compared results for the confidence score and the distance correspond to acceptances

wherein the likelihood of the context-dependent phonemes is calculated by;

log likelihood of context-dependent phoneme={(log likelihood of current phoneme)−

(mean of base phonemes of current phoneme)}/(standard deviation of base phonemes of current phoneme).

10. The utterance verification apparatus of claim 9, wherein the N-best utterance verifier accepts the N-best speech-recognized word when the confidence score is greater than the threshold, and rejects the N-best speech-recognized words when the confidence score is less than the threshold.

11. The utterance verification apparatus of claim 9, wherein the N-best utterance verifier accepts the N-best speech-recognized words when the distance is less than the mean of distances, and rejects the N-best speech-recognized words when the distance is greater than the mean of distances.

12. The utterance verification apparatus of claim 9, wherein the N-best utterance verifier rejects the N-best speech-recognized word, when the compared results for the confidence score and the distance correspond to rejections.

13. The utterance verification apparatus of claim 12, wherein the N-best utterance verifier determines indetermination for the N-best speech-recognized word, when any one of the compared results for the confidence score and the distance corresponds to acceptance and the other corresponds to rejection.

14. The utterance verification apparatus of claim 9, wherein the confidence score of the N-best speech-recognized word is calculated by:

word-level confidence score={(log likelihood of context-dependent word)−

(log likelihood of anti-phoneme model word)}/(the number of phonemes in word).

15. The utterance verification apparatus of claim 9, wherein the mean of distances is calculated by:

mean of distances=(sum of distances of N-best recognition results)/the number of N-best.

16. An utterance verification system for an isolated word N-best speech recognition result comprising:

a computer comprising;

a pre-processor extracting a feature vector of an input utterance and performing endpoint detection;

an N-best speech recognizer performing N-best speech recognition through Viterbi search by referring to the context-dependent phoneme model extracted from the feature vector; and

an N-best utterance verifier calculating log likelihoods of a context-dependent phoneme and an anti-phoneme model for the N-best speech-recognized word, comparing a confidence score measured for the N-best speech-recognized word with a threshold, comparing a distance measured for the N-best speech-recognized-word with a mean of distances, and accepting the N-best speech-recognized word when the compared results for the confidence score and the distance correspond to acceptances

wherein the likelihood of the anti-phoneme model is calculated by;

likelihood of anti-phoneme model={(log likelihood of anti-phoneme model of current phoneme)−

(mean of anti-phoneme model of current phoneme)}/(standard deviation of anti-phoneme model of current phoneme).