Speech analysis method and apparatus

• US 5,345,535 A
• Filed: 07/14/1993
• Issued: 09/06/1994
• Est. Priority Date: 04/04/1990
• Status: Expired due to Term

• Current Assignee: George R. Doddington
• Original Assignee: George R. Doddington
• Inventors: Doddington, George R.
• Primary Examiner(s): Knepper, David D.
• Application Number:
  US08/092,654
• Time in Patent Office: 419 Days
• Field of Search: 395/2, 395/2.4, 395/2.45-2.5, 381/41-50
• US Class Current: 704/236
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates

Abstract

System (100) receives a speech signal at an input (102) which is measured and transformed by speech feature measuring device (104). The output feature vector from speech feature measuring device (104) is then compared to a reference model in a statistical classification manner. Acoustic similarity measuring device (106) performs statistical measurements while temporal speech model constraints block (108) imposes transitional probabilities to the probability measurements generated by measuring device (106). Acoustic similarity measuring device (106) performs a weighted analysis of the error vector defined between the speech feature vector and reference vector utilized during the analysis.

Claims

1. A method of speech analysis comprising:

receiving a speech signal;

converting the speech signal into a converted signal for processing by a processor and memory system;

generating a feature vector from said converted signal, said feature vector having a plurality of feature vector elements;

providing a reference model comprising a plurality of states, each of said states comprising an associated means vector and covariance matrix;

generating an error vector having a plurality of error elements, each of said error elements corresponding to one of said feature vector elements;

weighting each of said error elements by a respective weight factor raised by an exponent, each respective weight factor comprising a factor proportional to a relative variance of each of said feature vector elements;

generating an observation score based on said feature vector, said weighted error elements, and said reference model states; and

based on a series of said observation scores, determining the probability that received speech signals correspond to a particular series of said reference model states.

2. The method of claim 1 wherein said step of weighting said error elements comprises weighting only said error elements corresponding to said feature vector elements having variances below a predetermined value.

3. The method of claim 1 wherein the error vector is defined as a product of a diagonal matrix of scale factors, wherein each scale factor is a reciprocal of a square root of an eigenvalue of said covariance matrix, times an eigenvector matrix associated with said reference model, times a difference between the feature vector and said mean vector.

4. The method of claim 1, wherein said exponent is in a range of zero to negative one.

5. The method of claim 1 wherein said exponent is negative one-half.

6. The method of claim 1 wherein said weighting step yields a plurality of partial results, and said step of generating an observation score further comprises:

squaring the partial results;

summing the squared partial results;

multiplying the summed, squared partial results times negative one-half; and

adding a covariance dispersion factor to the multiplied, summed, squared partial results.

7. An apparatus for performing speech analysis, comprising:

circuitry for receiving a speech signal;

circuitry for converting said speech signal to a converted signal for processing by a processor and memory system;

circuitry for transmitting said converted signal to speech feature measuring circuitry;

said speech feature measuring circuitry for generating a feature vector from said converted signal, said feature vector having a plurality of feature vector elements;

a memory for storing a reference model comprising a plurality of states, each of said states comprising an associated mean vector and covariance matrix;

acoustic similarity measuring circuitry, including;

circuitry for generating an error vector having a plurality of error elements, each of said error elements corresponding to one of said feature vector elements;

circuitry for weighting each of said error elements of the error vector by a respective weight factor raised by an exponent, each respective weight factor comprising a factor proportional to a relative variance of each of said feature vector elements;

circuitry for generating an observation score based on said feature vector, said weighted error elements, and said reference model states; and

circuitry for determining, based on a series of said observation scores, the probability that received speech signals correspond to a particular series of said reference model states.

8. The apparatus of claim 7 wherein said circuitry for weighting said error elements comprises circuitry for weighting only said error elements corresponding to said feature vector elements having variances below a predetermined value.

9. The apparatus of claim 7 wherein said error vector is defined as a diagonal matrix of scale factors, wherein each scale factor is a reciprocal of a square root of an eigenvalue of said covariance matrix, times selected elements of an eigenvector matrix associated with said reference model, time corresponding selected elements of a difference between said feature vector and said mean vector.

10. The apparatus of claim 7 wherein said exponent is in a range of zero to negative one.

11. The apparatus of claim 7 wherein said exponent is negative one-half.