Cepstral correction vector quantizer for speech recognition

• US 5,598,505 A
• Filed: 09/30/1994
• Issued: 01/28/1997
• Est. Priority Date: 09/30/1994
• Status: Expired due to Term

• Current Assignee: Apple Inc.
• Original Assignee: Apple Computer Incorporated (Apple Inc.)
• Inventors: Austin, Stephen C., Fineberg, Adam B.
• Primary Examiner(s): Tung, Kee M.
• Application Number:
  US08/316,118
• Time in Patent Office: 851 Days
• Field of Search: 395/2.3-2.33, 395/2.25-2.28, 395/2.35, 395/2.6-2.66, 395/2.39, 381/46-47, 381/34, 381/35, 381/43, 381/41
• US Class Current: 704/226
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering
• CPC Class Codes: G10L 2019/0004 Design or structure of the codebook
• CPC Class Codes: G10L 25/24 the extracted parameters being the cepstrum

Abstract

A method for correcting cepstral vectors representative of speech generated in a test environment by use of a vector quantization (VQ) system with a codebook of vectors that was generated using speech and acoustic data from a different (training) environment. The method uses a two-step correction to produce test environment cepstral vectors with reduced non-speech acoustic content. The first correction step subtracts, from the test vector, a coarse correction vector that is computed from an average of test environment cepstral vectors. The second step involves a VQ of the coarsely corrected test vector at each node of the VQ tree. The third step is the addition of a fine correction vector to the coarsely corrected test vector that is generated by subtracting a running (moving) average of the coarsely corrected test vectors associated with the deepest VQ tree node from the VQ vector closest to the coarsely corrected test vector. The method is independent of the means used to generate the cepstral vectors and the corrected output cepstra vectors may be used in various speech processing and classifying systems. The method is adaptable to non-stationary environments.

Claims

1. A method for correcting a cepstral test vector representation of speech from an acoustical test environment using a vector quantizer (VQ) encoder having a VQ encoder codebook based on training data from a different acoustical training environment, the method comprising:

(a) applying a coarse correction vector from the cepstral test vector, the coarse correction vector being representative of the acoustical training environment without the presence of speech, for producing a coarsely corrected cepstral test vector; and

(b) applying a fine correction vector to the coarsely corrected cepstral vector for producing a fine corrected cepstral test vector, the fine correction vector representative of a difference between acoustical test environment with the presence of speech only and the acoustical training environment cepstral training vectors with the presence of speech only.

2. The method of claim 1 wherein the coarse correction vector is an average of a set of cepstral test vectors generated in the acoustic test environment.

3. The method of claim 1 wherein the fine correction vector is representative of a difference between a VQ encoder codebook vector that is closest to a coarsely corrected cepstral test vector and a vector representative of a running average of test cepstral vectors that is closest to the VQ encoder codebook vector.

4. A method for correcting cepstral vectors representative of speech from an acoustical test environment for use in a speech processing system by using a vector quantization (VQ) codebook based on training data from an acoustical training environment, and by applying correction vectors to the cepstral vectors, the method comprising:

(a) acquiring a set of cepstral vectors representative of speech signals from the acoustical training environment;

(b) computing an average cepstral vector from the acquired set of cepstral vectors, the average cepstral vector representative of the acoustical training environment without speech;

(c) correcting the set of cepstral vectors by subtracting the average cepstral vector of step (b) from each of the acquired set of cepstral vectors of step (a) for producing a set of coarsely corrected vectors;

(d) generating a VQ codebook tree from the set of coarsely corrected vectors of step (c) in which each codebook tree node is a vector that is representative of a set of coordinates describing a centroid of a distinct cluster, each cluster having an assigned subset of the set of coarsely corrected vectors of step (c), the subset of coarsely corrected vectors in each cluster being closer to the centroid of the cluster to which the subset is assigned than to any other cluster;

(e) acquiring new cepstral vectors from a test environment and computing a running average test vector,

(f) correcting each new cepstral vector of from the test environment by

(i) subtracting the running average cepstral vector from step (e) from each new cepstral vector to obtain a coarsely corrected new vector,

(ii) vector quantizing the coarsely corrected new vector of step (f)(i) using the VQ codebook tree of step (d) to obtain a VQ vector,

(iii) accumulating a running average node vector from a set of new vectors associated with a node of the VQ codebook tree to which each vector of the set of new vectors is closest,

(iv) generating a fine correction vector from the VQ vector of step (f)(ii) by subtracting the running average node vector corresponding to the VQ vector node,

(v) producing a finely corrected new vector by adding the fine correction vector to the coarsely corrected new vector; and

(g) outputting the finely corrected new vector for processing by the speech processing system.

5. The method of claim 4, further comprising an adaptation method for correcting cepstral vectors, the adaptation method comprising:

(h) computing a new average cepstral vector by including in the average cepstral vector of step (b) the new cepstral vector; and

(j) computing a new centroid for the cluster associated with the VQ vector of step (e)(ii) by including within the cluster the coarsely corrected new vector.

6. The method of claim 5, wherein both step (h) and step (j) use moving average techniques for computing the average cepstral vector and for computing the new centroid.

7. The method of claim 6 wherein the moving average technique is an exponentially discounted moving average technique.

8. A method for use in speech processing system for pre-processing a cepstral test vector that is representative of a speech signal in a test acoustical environment that can include extraneous quasi-stationary non-speech acoustic signals, the method reduces the extraneous signals by subtracting average cepstral vectors generated in a training acoustic environment and the test acoustical environment, the method comprises:

(a) generating correction means using a set of training cepstral vectors, including,

(i) generating an average training vector from the set of training cepstral vectors,

(ii) producing a set of mean normalized training vectors by subtracting the mean value corresponding to the average training vector,

(iii) producing a vector quantization (VQ) codebook tree based on the set of mean normalized training vectors; and

(b) acquiring and correcting each vector of a sequence of cepstral test vectors, including,

(i) generating a running mean cepstral test vector from the sequence of cepstral test vectors,

producing a coarsely corrected test vector by subtracting the running mean cepstral test vector from each cepstral test vector,

(iii) vector quantizing the coarsely corrected test vector using the VQ codebook tree of step (a)(iii) and producing a VQ output vector corresponding to a vector quantized coarsely corrected test vector,

(iv) computing a running average of coarsely corrected test vectors associated with each closest VQ codebook tree node, and

(v) producing an output vector with reduced extraneous signals by adjusting the VQ output vector of step (b)(iii) by adding a fine correction vector formed by subtracting the running average coarsely corrected cepstral test vector from the vector quantized coarsely corrected test vector.

9. The method of claim 8 wherein the VQ codebook of step (a)(iii) is based on a binary tree.

10. The method of claim 8 wherein the running mean cepstral test vector of step (b)(i) is computed using an exponentially discounted moving average algorithm.