Method and apparatus for utterance verification

• US 8,972,264 B2
• Filed: 12/17/2012
• Issued: 03/03/2015
• Est. Priority Date: 11/08/2012
• Status: Active Grant

• Current Assignee: Industrial Technology Research Institute
• Original Assignee: Industrial Technology Research Institute
• Inventors: Chien, Shih-Chieh
• Primary Examiner(s): Abebe, Daniel D
• Application Number:
  US13/717,645
• Publication Number:
  US 20140129224A1
• Time in Patent Office: 806 Days
• Field of Search: 704/251, 704/255, 704/257
• US Class Current: 704/251
• CPC Class Codes: G10L 15/01 Assessment or evaluation of speech recognition systems
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 2015/085 Methods for reducing search complexity, pruning

Abstract

A method and apparatus for utterance verification are provided for verifying a recognized vocabulary output from speech recognition. The apparatus for utterance verification includes a reference score accumulator, a verification score generator and a decision device. A log-likelihood score obtained from speech recognition is processed by taking a logarithm of the value of the probability of one of feature vectors of an input speech conditioned on one of states of each model vocabulary. A verification score is generated based on the processed result. The verification score is compared with a predetermined threshold value so as to reject or accept the recognized vocabulary.

Claims

1. A method for utterance verification adapted to verify a recognized vocabulary, wherein the recognized vocabulary is obtained by performing speech recognition on a feature vector sequence according to an acoustic model and model vocabulary database, wherein the feature vector sequence comprises feature vectors of a plurality of frames, wherein the acoustic model and model vocabulary database comprises a plurality of model vocabularies, wherein each of the model vocabularies comprises a plurality of states, and wherein the method for utterance verification comprises:

calculating a maximum reference score for each of the model vocabularies according to a log-likelihood score obtained from speech recognition, wherein the log-likelihood score obtained from speech recognition is calculated by taking a logarithm on a value of a probability of one of the feature vectors of the frames conditioned on one of the states of each model vocabulary, and wherein the maximum reference score is a summation of the maximum value of log-likelihood scores of the feature vector of each frame conditioned on each state of a certain model vocabulary;

calculating a first verification score according to an optimal path score output during the speech recognition and the maximum reference score; and

comparing the first verification score with a first predetermined threshold value, so as to reject or accept the recognized vocabulary.

2. The method for utterance verification as claimed in claim 1, wherein an equation for calculating the first verification score is:

3. A method for utterance verification adapted to verify a recognized vocabulary, wherein the recognized vocabulary is obtained by performing speech recognition on a feature vector sequence according to an acoustic model and model vocabulary database, wherein the feature vector sequence comprises feature vectors of a plurality of frames, wherein the acoustic model and model vocabulary database comprises a plurality of model vocabularies, wherein each of the model vocabularies comprises a plurality of states, and wherein the method for utterance verification comprises:

calculating an overall maximum reference score according to a log-likelihood score obtained from speech recognition, wherein the log-likelihood score obtained from speech recognition is calculated by taking a logarithm on a value of a probability of one of the feature vectors of the frames conditioned on one of the states of each model vocabulary, and wherein the overall maximum reference score is a summation of the maximum value of log-likelihood scores of the feature vector of each frame conditioned on each state of each of the model vocabularies;

calculating a second verification score according to an optimal path score output during the speech recognition and the overall maximum reference score; and

comparing the second verification score with a second predetermined threshold value, so as to reject or accept the recognized vocabulary.

4. The method for utterance verification as claimed in claim 3, wherein an equation for calculating the second verification score is:

5. The method for utterance verification as claimed in claim 3 further comprising:

calculating a garbage score according to a garbage model, wherein the garbage score is obtained by taking a logarithm on a value of a probability of one of the feature vectors conditioned on a state of the garbage model;

calculating a third verification score according to the optimal path score, the garbage score and the overall maximum reference score; and

comparing the third verification score with a third predetermined threshold value, so as to reject or accept the recognized vocabulary.

6. The method for utterance verification as claimed in claim 5, wherein an equation for calculating the third verification score is:

7. The method for utterance verification as claimed in claim 3 further comprising:

calculating an overall minimum reference score, wherein the overall minimum reference is a summation of the minimum value of log-likelihood scores of the feature vector of each frame conditioned on each state of each of the model vocabularies;

calculating a fourth verification score according to the optimal path score, the overall maximum reference score and the overall minimum reference score; and

comparing the fourth verification score with a fourth predetermined threshold value, so as to reject or accept the recognized vocabulary.

8. The method for utterance verification as claimed in claim 7, wherein an equation for calculating the fourth verification score is:

9. An apparatus for utterance verification adapted to verify a recognized vocabulary output by a speech recognition device, wherein the recognized vocabulary is obtained by performing speech recognition on a feature vector sequence according to an acoustic model and model vocabulary database, wherein the feature vector sequence comprises feature vectors of a plurality of frames, wherein the acoustic model and model vocabulary database comprises a plurality of model vocabularies, wherein each of the model vocabularies comprises a plurality of states, and wherein the apparatus for utterance verification comprises:

a reference score accumulator coupled to the speech recognition device and adapted to calculate a maximum reference score for each of the model vocabularies according to a log-likelihood score obtained from the speech recognition device by taking a logarithm on a value of a probability of one of the feature vectors of the frames conditioned on one of the states of each model vocabulary, wherein the maximum reference score is a summation of the maximum value of log-likelihood scores of the feature vector of each frame conditioned on each state of a certain model vocabulary;

a verification score generator coupled to the reference score accumulator and adapted to calculate a first verification score according to an optimal path score output from the speech recognition device and the maximum reference score; and

a decision device coupled to the verification score generator and adapted to compare the first verification score with a first predetermined threshold value, so as to reject or accept the recognized vocabulary.

10. The apparatus for utterance verification as claimed in claim 9, wherein an equation for calculating the first verification score is:

11. An apparatus for utterance verification adapted to verify a recognized vocabulary output by a speech recognition device, wherein the recognized vocabulary is obtained by performing speech recognition on a feature vector sequence according to an acoustic model and model vocabulary database, wherein the feature vector sequence comprises feature vectors of a plurality of frames, wherein the acoustic model and model vocabulary database comprises a plurality of model vocabularies, wherein each of the model vocabularies comprises a plurality of states, and wherein the apparatus for utterance verification comprises:

a reference score accumulator coupled to the speech recognition device and adapted to calculate an overall maximum reference score according to a log-likelihood score obtained from the speech recognition device by taking a logarithm on a value of a probability of one of the feature vectors of the frames conditioned on one of the states of each model vocabulary, wherein the overall maximum reference score is a summation of the maximum value of log-likelihood scores of the feature vector of each frame conditioned on each state of each of the model vocabularies;

a decision device coupled to the reference score accumulator and adapted to calculate a second verification score according to an optimal path score output by the speech recognition device and the overall maximum reference score;

a verification score generator coupled to the reference score accumulator and adapted to compare the second verification score with a second predetermined threshold value, so as to reject or accept the recognized vocabulary.

12. The apparatus for utterance verification as claimed in claim 11, wherein an equation for calculating the second verification score is:

13. The apparatus for utterance verification as claimed in claim 11, wherein the reference score accumulator further calculates a garbage score according to a garbage model, wherein the garbage score is obtained by taking a logarithm on a value of a probability of one of the feature vectors conditioned on a state of the garbage model, wherein the verification score generator calculates a third verification score according to the optimal path score, the garbage score and the overall maximum reference score, and wherein the decision device compares the third verification score with a third predetermined threshold value so as to reject or accept the recognized vocabulary.

14. The apparatus for utterance verification as claimed in claim 13, wherein an equation for calculating the third verification score is:

15. The apparatus for utterance verification as claimed in claim 11, wherein the reference score accumulator further calculates an overall minimum reference score, wherein the overall minimum reference score is a summation of the minimum value of log-likelihood scores of the feature vector of each frame conditioned on each state of each of the model vocabularies, wherein the verification score generator calculates a fourth verification score according to the optimal path score, the overall maximum reference score and the overall minimum reference score, and wherein the decision device compares the fourth verification score with a fourth predetermined threshold value so as to reject or accept the recognized vocabulary.

16. The apparatus for utterance verification as claimed in claim 15, wherein an equation for calculating the fourth verification score is: