Method and system for continuous speech recognition using voting techniques

• US 5,638,486 A
• Filed: 10/26/1994
• Issued: 06/10/1997
• Est. Priority Date: 10/26/1994
• Status: Expired due to Term

• Current Assignee: Google Technology Holdings LLC (Alphabet Inc.)
• Original Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Inventors: Lindsey, Michael K., Wang, Shay-Ping T.
• Primary Examiner(s): Knepper, David D.
• Application Number:
  US08/329,394
• Time in Patent Office: 958 Days
• Field of Search: 395/2, 395/2.12, 395/2.13, 395/2.26, 395/2.45, 395/2.57, 395/2.62, 395/2.63, 395/2.41, 381/41-43
• US Class Current: 704/236
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates

Abstract

In a speech-recognition system having a plurality of classifiers, a voting window includes a sequence of outputs from each of the classifiers. For each classifier, a voting sum is generated corresponding to the voting window. A spoken sound is identified by determining which classifier corresponds to the greatest voting sum.

Claims

1. In a speech-recognition system having a plurality of classifiers, a method of identifying a spoken sound, comprising the following steps:

(a) receiving a plurality of classifier output signals from the classifiers corresponding to an interval, each of the classifier output signals having been generated according to a polynomial discriminant function;

(b) ranking by magnitude the classifier output signals to produce a rank-order of classifier output signals corresponding to the interval;

(c) weighting each position in the rank-order to transform the classifier output signals into a plurality of weighted values;

(d) repeating steps (a)-(c) for a plurality of intervals, whereby generating a plurality of weighted value sequences, each of the weighted value sequences corresponding to a respective one of the plurality of classifiers;

(e) summing each of the weighted value sequences to generate a voting sum for each classifier; and

(f) identifying the spoken sound by selecting the voting sum having a largest magnitude.

2. The method of claim 1 further comprising the step of:

generating a system output which includes a class label representing the spoken sound.

3. The method of claim 1 wherein the polynomial discriminant function has a form ##EQU3## wherein x_j represents a plurality of features;

wherein i, j, m and n are integers, y represents a classifier output signal;

wherein w_i represents a coefficient;

wherein g_ji represents an exponent.

4. The method of claim 3, wherein the polynomial discriminant function has the form ##EQU4## wherein a₀ represents a zero-order coefficient, b_i represents a first-order coefficient, and c_ij represents a second-order coefficient.

5. The method of claim 1 wherein the speech-recognition system identifies a plurality of spoken sounds from continuous speech.

6. The method of claim 1 wherein steps (a)-(c) are repeated for three successive intervals.

7. The method of claim 1 wherein the spoken sound is selected from the group consisting of word, syllable, and phoneme.

8. A method for recognizing a spoken sound from continuous speech, comprising the following steps:

(a) receiving the continuous speech;

(b) sampling the continuous speech, over time, to form a sequence of sample datum which represents the continuous speech;

(c) partitioning the sequence of sample datum into a sequence of data frames, each of the sequence of data frames includes at least two of the sequence of sample datum;

(d) extracting a plurality of features from the sequence of data frames;

(f) forming a sequence of feature frames from the plurality of features;

(g) distributing one of the sequence of feature frames to a plurality of classifiers, each of the classifiers generating a classifier output signal in response thereto according to a polynomial discriminant function, whereby producing a plurality of classifier output signals;

(h) ranking by magnitude the classifier output signals to produce a rank-order of classifier output signals corresponding to the distributed feature frame;

(i) weighting each position in the rank-order to transform the classifier output signals into a plurality of weighted values;

(j) repeating steps (g)-(i) for each feature frame included in the sequence of feature frames, whereby generating a plurality of weighted value sequences, each of the weighted value sequences corresponding to a respective one of the classifiers;

(k) summing each of the weighted value sequences to generate a voting sum for each classifier; and

(l) identifying the spoken sound by selecting the voting sum having a largest magnitude.

9. The method of claim 8, further comprising the step of:

generating a system output which includes a class label representing the spoken sound.

10. The method of claim 8 wherein the polynomial discriminant function has a form ##EQU5## wherein x_j represents the features included in the distributed feature frame;

wherein i, j, m and n are integers, wherein y represents the classifier output signal;

wherein w_i represents a coefficient;

wherein g_ji represents an exponent.

11. The method of claim 10, wherein the polynomial discriminant function has the form ##EQU6## wherein a₀ represents a zero-order coefficient, b_i represents a first-order coefficient, and c_ij represents a second-order coefficient.

12. The method of claim 8 wherein the speech-recognition system recognizes a plurality of spoken sounds from the continuous speech.

13. The method of claim 8 wherein the sequence of feature frames consists of three feature frames.

14. The method of claim 8 wherein the spoken sound is selected from the group consisting of word, syllable, and phoneme.

15. A speech-recognition system for identifying a spoken sound and having a plurality of classifiers, comprising:

receiving means for receiving a plurality of classifier output signals from the classifiers corresponding to an interval, each of the classifier output signals having been generated according to a polynomial discriminant function;

ranking means, associatively coupled to the receiving means, for ranking by magnitude the classifier output signals to produce a rank-order of classifier output signals corresponding to the interval;

weighting means, associatively coupled to the ranking means, for weighting each position in the rank-order to transform the classifier output signals into a plurality of weighted values;

summing means, associatively coupled to the weighting means, for respectively summing a plurality of weighted value sequences to generate a plurality of voting sums, each of the voting sums corresponding to a respective one of the plurality of classifiers; and

identifying means, associatively coupled to the summing means, for identifying the spoken sound by selecting from the plurality of voting sums a voting sum having a largest magnitude the subsequence to produce a voting sum for each of the plurality of classifiers,

wherein the receiving means, the defining means, and the weighting means cooperatively function over a plurality of intervals to generate the plurality of weighted value sequences.

16. The speech-recognition system of claim 15 wherein the identifying means generates a system output which includes a class label representing the spoken sound.

17. The speech-recognition system of claim 15, wherein the polynomial discriminant function has a form ##EQU7## wherein x_j represents a plurality of features;

wherein i, j, m and n are integers, y represents a classifier output signal;

wherein w_i represents a coefficient;

wherein g_ji represents an exponent.

18. The method of claim 17, wherein the polynomial discriminant function has the form ##EQU8## wherein a₀ represents a zero-order coefficient, b_i represents a first-order coefficient, and c_ij represents a second-order coefficient.

19. The speech-recognition system of claim 15 wherein the speech-recognition system identifies a plurality of spoken sounds from continuous speech.

20. The speech-recognition system of claim 15 wherein the subsequence includes a sequence of three outputs from one of the plurality of classifiers.

21. The speech-recognition system of claim 15 wherein the spoken sound is selected from the group consisting of word, syllable, and phoneme.