Speech recognition method using speaker cluster models

• US 6,567,776 B1
• Filed: 04/04/2000
• Issued: 05/20/2003
• Est. Priority Date: 08/11/1999
• Status: Expired due to Term

• Current Assignee: Industrial Technology Research Institute
• Original Assignee: Industrial Technology Research Institute
• Inventors: Chien, Shih-Chieh, Chang, Sen-Chia, Penwu, Chung-Mou
• Primary Examiner(s): Banks-Harold, Marsha D.
• Assistant Examiner(s): Lerner, Martin
• Application Number:
  US09/542,844
• Time in Patent Office: 1,141 Days
• Field of Search: 704/236, 704/240, 704/243, 704/244, 704/245, 704/246, 704/250
• US Class Current: 704/236
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

In speaker-independent speech recognition, between-speaker variability is one of the major resources of recognition errors. A speaker cluster model is used to manage recognition problems caused by between-speaker variability. In the training phase, the score function is used as a discriminative function. The parameters of at least two cluster-dependent models are adjusted through a discriminative training method to improve performance of the speech recognition.

Claims

1. A speech recognition method comprising:

receiving a speech signal;

recognizing the speech signal using a speaker cluster model obtained in a training phase wherein the speaker cluster model is a collection of a plurality of cluster-dependent models, and a score of each candidate is calculated according to a score function which is defined by taking the dependency among the cluster-dependent models into account; and

obtaining a final recognition result according to a decision rule based on the Score of each candidate,

wherein the training phase comprises building an initialization model, and adjusting parameters of at least two cluster-dependent models of the initialization model by using a discriminative training method to obtain the speaker cluster model wherein the discriminative training method is implemented by using a minimum classification error as a training criterion, a discriminant function of the discriminative training method being defined in the same manner as the score function, and

the score function is defined as;

$g_i<br><br>\left(X;<br><br>\mathrm{\Gamma <br><br>}\right)={\log <br><br>\text{\hspace{1em}<br><br>}\left[\frac{1}{N}<br><br>\underset{n=1}{\overset{N}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>w_n<br><br>\left(X\right)<br><br>\text{\hspace{1em}<br><br>}<br><br>\exp <br><br>\text{\hspace{1em}<br><br>}\left[h_i<br><br>\left(X;<br><br>{\mathrm{\Lambda <br><br>}}_n\right)<br><br>\mathrm{\xi <br><br>}\right]\right]}^{\frac{1}{\mathrm{\xi <br><br>}}},i=1,2,\mathrm{\cdots <br><br>}<br><br>\text{\hspace{1em}<br><br>},M$

 

wherein g_i(X;

Ã

) is the score function, X is a feature vector sequence of the speech signal, Ã

represents an entire parameter set of the speaker cluster model, N is the number of cluster-dependent models, parameter sets corresponding to the N cluster-dependent models are Ë

₁, Ë

₂, . . . , Ë

_N, M is the number of candidates to be classified, h_i(X;

Ë

_n) is a log-likelihood function defined only on a parameter set Ë

_n, î

is a positive weighting number, and w_n(X) is a cluster weighting function that indicates the degree to which the nth cluster-dependent model is used for recognition.

2. The speech recognition method of claim 1 wherein the discriminative training method is implemented by a generalized probabilistic descent method.

3. The speech recognition method of claim 1 wherein the decision rule is to selected a candidate with a highest score.

4. The speech recognition method of claim 1 wherein the cluster weighting function w_n(X) is a zero-one function used to indicate whether the nth cluster-dependent model will be used for recognition.