Discriminative clustering methods for automatic speech recognition

• US 6,526,379 B1
• Filed: 11/29/1999
• Issued: 02/25/2003
• Est. Priority Date: 11/29/1999
• Status: Expired due to Term

• Current Assignee: Sovereign Peak Ventures, LLC (Dominion Harbor Enterprises, LLC)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Tsakam, Brice, Rigazio, Luca, Junqua, Jean-Claude
• Primary Examiner(s): CHAWAN, VIJAY B
• Application Number:
  US09/450,387
• Time in Patent Office: 1,184 Days
• Field of Search: 704/245, 704/256, 704/255, 704/240, 704/243, 704/244, 704/236, 704/246, 704/222, 704/231, 704/238
• US Class Current: 704/245
• CPC Class Codes: G10L 15/06 Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice G10L15/14 takes precedence
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

The discriminative clustering technique tests a provided set of Gaussian distributions corresponding to an acoustic vector space. A distance metric, such as the Bhattacharyya distance, is used to assess which distributions are sufficiently proximal to be merged into a new distribution. Merging is accomplished by computing the centroid of the new distribution by minimizing the Bhattacharyya distance between the parameters of the Gaussian distributions being merged.

Claims

1. A discriminative clustering method for automatic speech recognition, comprising:

providing a set of Gaussian distributions corresponding an acoustic vector space;

testing said distributions to identify selected distributions having a predetermined proximity to one another based on a distance measure and merging said selected distributions into at least one new distribution;

computing the centroid of said new distribution by minimizing the Bhattacharyya distance between selected distribution and its centroid.

2. The method of claim 1 wherein said distance measure is the Bhattacharyya distance.

3. The method of claim 1 further comprising organizing said set of Gaussian distributions into a tree structure and wherein said testing said distributions is performed following said tree structure.

4. The method of claim 1 wherein said merging step is performed by selecting a couple of distributions, computing an overall distortion resulting from the merging of said couple of distributions and committing said merging if that results in a lower overall distortion.

5. The method of claim 1 wherein said merging step is iteratively performed by selecting a couple of distributions, computing an overall distortion resulting from the merging of said couple of distributions and committing said merging if that results in a lower overall distortion.

6. The method of claim 1 further comprising performing a Lloyd-Max process on said new distribution.

7. A method for automatic speech recognition, comprising:

providing a set of Gaussian distributions corresponding an acoustic vector space;

a step for testing said distributions to identify selected distributions;

a step for merging said selected distributions into at least one new distribution; and

a step for computing the centroid of said new distribution.

8. The method of claim 7 further comprising a step for organizing said set of Gaussian distributions into a tree structure following said step for testing said distributions.

9. The method of claim 7 further comprising a step for iteratively performing said step for merging said selected distributions.

10. A discriminative clustering method for automatic speech recognition, comprising:

providing a set of Gaussian distributions corresponding an acoustic vector space;

computing a Bhattacharyya distance measure between each of said distributions;

selecting distributions having a predetermined proximity to one another based on a distance measure; and

merging said selected distributions into at least one new distribution.

11. The method of claim 10 wherein the Bhattacharyya distance between two Gaussian distributions g₁ and g₂ is computed in accordance with $D_{\mathrm{bhat}}$

(g1,g2)=18*(μ

2-μ

1)T*[∑

1

+∑

22]-1*(μ

2-μ

1)+12

ln



∑

1

+∑

2

2

∑

1

*

∑

2



 

where μ

₂ and μ

₂ are the mean vectors associated with distributions g₁ and g₂ and Σ

₁ and Σ

₂ are the respective covariance matrices.

12. The method of claim 10 further comprising:

computing a centroid for said new distribution by minimizing the Bhattacharyya distance between said selected distributions.

13. The method of claim 10 wherein the centroid of the new distribution which contains Gaussian distributions g_i with a weight of w_i, i=1, . . . , N is computed in accordance with ${\mu }_{}$

c=∑

i

wi*(σ

c2+σ

i2)-1*μ

i∑

i

wi*(σ

c2+σ

i2)-1${\mathrm{\sigma <br><br>}}_c^2=\frac{{\sum <br><br>}_i<br><br>w_i}{{\sum <br><br>}_i<br><br>w_i*\left[\frac{2}{{\mathrm{\sigma <br><br>}}_c^2+{\mathrm{\sigma <br><br>}}_i^2}-{\left(\frac{{\mathrm{\mu <br><br>}}_c-{\mathrm{\mu <br><br>}}_i}{{\mathrm{\sigma <br><br>}}_c^2+{\mathrm{\sigma <br><br>}}_i^2}\right)}^2\right]}$

where σ

² is the variance and μ

is the mean associated with a specific component of the feature vector.