Discriminative utterance verification for connected digits recognition

• US 5,737,489 A
• Filed: 09/15/1995
• Issued: 04/07/1998
• Est. Priority Date: 09/15/1995
• Status: Expired due to Fees

• Current Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Rahim, Mazin G., Chou, Wu, Juang, Biing-Hwang, Lee, Chin-Hui
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): COLLINS, ALPHONSO
• Application Number:
  US08/528,902
• Time in Patent Office: 935 Days
• Field of Search: 395/2.4, 395/2.41, 395/2.45, 395/2.47, 395/2.48, 395/2.49, 395/2.51, 395/2.6, 395/2.65
• US Class Current: 704/256
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates
• CPC Class Codes: G10L 15/144 Training of HMMs
• CPC Class Codes: G10L 2015/225 Feedback of the input speech

Abstract

In a speech recognition system, a recognition processor receives an unknown utterance signal as input. The recognition processor in response to the unknown utterance signal input accesses a recognition database and scores the utterance signal against recognition models in the recognition database to classify the unknown utterance and to generate a hypothesis speech signal. A verification processor receives the hypothesis speech signal as input to be verified. The verification processor accesses a verification database to test the hypothesis speech signal against verification models reflecting a preselected type of training stored in the verification database. Based on the verification test, the verification processor generates a confidence measure signal. The confidence measure signal can be compared against a verification threshold to determine the accuracy of the recognition decision made by the recognition processor.

Claims

1. A speech signal processing method of making a speech recognizer verification model database based on one or more known hypothesis string signals representing known speech utterances and a set of current verification models, the method comprising the following steps:

(A) receiving a known hypothesis string signal representing a known speech utterance;

(B) generating a first string-based confidence measure signal based on the known hypothesis string signal representing a known speech utterance and a current verification model for that signal;

(C) generating one or more other string-based confidence measure signals, each such string-based confidence measure signal based on the known hypothesis string signal representing a known speech utterance and another current verification model;

(D) computing a misverification signal based on the first string-based confidence measure signal and the other string-based confidence measure signals;

(E) based on the misverification signal and the known hypothesis string signal representing a known speech utterance, modifying one or more of the current verification models to decrease the likelihood of misverifying an unknown hypothesis string signal representing an unknown speech utterance; and

(F) storing one or more modified verification models in memory.

2. A method according to claim 1, further comprising the step:

initializing the verification models in the verification database.

3. A method according to claim 1, further comprising the steps:

computing a loss function signal based on the misverification signal; and

determining a gradient of the loss function signal including calculating the derivative of the loss function signal.

4. A method according to claim 1, wherein step (B) comprises the steps:

segmenting the known hypothesis string signal into a series of word signals;

calculating a word-based confidence score for each word signal of the series; and

averaging the word-based confidence scores calculated for each word signal.

5. A method according to claim 1, wherein step (C) comprises the steps:

generating a set of confusable string models based on the known hypothesis string signal, a confusable string model comprising one or more word models;

generating a string-based confidence measure signal for each of the set of confusable string models; and

calculating an average of the string-based confidence measure signals for each of the set of confusable string models.

6. A method according to claim 5, wherein step (C) further comprises the steps:

determining a word-based confidence score for each word model of each of the set of confusable string models; and

averaging the word-based confidence scores determined for each word model of each of the set of confusable string model to generate the string-based confidence measure signal for each of the set of confusable string models.

7. A method according to claim 1, wherein step (D) comprises the step:

determining a gradient of a function relating

(a) confidence measure signal generation for the known hypothesis string signal based on a current verification model for that hypothesis string signal to

(b) confidence measure signal generation for the known hypothesis string signal based on one or more other current verification models.

8. A method according to claim 7, wherein step (E) comprises the step:

adjusting one or more parameters of the current verification models based on the gradient.

9. A method according to claim 1, wherein step (D) comprises the step:

forming a difference between

(a) the first string-based confidence measure signal and

(b) an average of the one or more other confidence measure signals.

10. A method according to claim 3, further comprising the steps:

estimating an expected loss function signal; and

iteratively adjusting the verification models responsive to the estimated loss function signal to minimize the likelihood of verification error.

11. A speech recognition system for recognizing a string of connected spoken words, comprising:

a recognition processor for generating a hypothesis string signal representing an unknown speech utterance responsive to an acoustic input string of connected spoken words;

a recognition database for storing recognition models;

a verification processor for generating a string-based confidence measure signal responsive to the hypothesis string signal representing an unknown speech utterance; and

a verification database for storing verification models.

12. A speech recognition system according to claim 11, further comprising:

means, coupled to the verification processor, for calculating a verification threshold signal value of a string-based confidence measure signal;

means for measuring the string-based confidence measure signal to generate a measured signal value; and

means for comparing the measured signal value to the threshold signal value.

13. A speech recognition system according to claim 11, wherein:

the verification models comprise a set of hidden Markov model parameters.

14. A speech recognition system according to claim 11, wherein:

the verification models are selected from the group consisting of keyword models, anti-keyword models, acoustic filler models, and combinations thereof.

15. A speech recognition system according to claim 11, wherein:

the verification models reflect discriminative training.

16. A speech recognition system according to claim 11, wherein:

the hypothesis string signal comprises a segmented series of word signals; and

the verification processor comprises

means for calculating a word-based confidence score for each word signal of the segmented series; and

means for generating the string-based confidence measure signal responsive to an average of the word-based confidence scores for each word signal of the segmented series.

17. A speech signal processing method for generating a verification threshold signal sample of a confidence measure signal generated by a verification processor in a speech recognition system, comprising the following steps:

(A) performing a round of operating the verification processor to

(i) generate a first string-based confidence measure signal based on a known hypothesis string signal representing a known speech utterance and a current verification model for the known hypothesis string signal representing a known speech utterance, and

(ii) generate one or more other string-based confidence measure signals, each such string-based confidence measure signal based on the known hypothesis string signal representing a known speech utterance and another verification model;

(B) repeating step (A) for a preselected number of rounds;

(C) recording a first distribution of first string-based confidence measure signal values for the preselected number of rounds;

(D) recording a second distribution of the one or more other string-based confidence measure signal values for the preselected number of rounds;

(E) calculating a first mean for the first distribution;

(F) calculating a second mean for the second distribution; and

(G) generating the verification threshold signal sample based on the first mean and the second mean.

18. A method according to claim 17, further comprising the steps:

selecting a false rejection verification error rate;

selecting a false acceptance verification error rate; and

estimating an initial verification threshold signal value based on the selected false rejection verification error rate and the selected false acceptance verification error rate.

19. A method according to claim 18, further comprising the steps:

generating the verification threshold signal value while the speech recognition system is operating; and

maintaining the false rejection verification error rate and the false acceptance verification error rate while the recognition system is operating.