Source normalization training for HMM modeling of speech

• US 6,151,573 A
• Filed: 08/15/1998
• Issued: 11/21/2000
• Est. Priority Date: 09/17/1997
• Status: Expired due to Term

• Current Assignee: Intel Corporation
• Original Assignee: Texas Instruments, Inc.
• Inventors: Gong, Yifan
• Primary Examiner(s): Isen, Forester W.
• Assistant Examiner(s): Azad, Abul K.
• Application Number:
  US09/134,775
• Time in Patent Office: 829 Days
• Field of Search: 704/236, 704/237, 704/239, 704/240, 704/243, 704/244, 704/255, 704/256, 704/233, 704/234
• US Class Current: 704/256.2
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

A maximum likelihood (ML) linear regression (LR) solution to environment normalization is provided where the environment is modeled as a hidden (non-observable) variable. By application of an expectation maximization algorithm and extension of Baum-Welch forward and backward variables (Steps 23a-23d) a source normalization is achieved such that it is not necessary to label a database in terms of environment such as speaker identity, channel, microphone and noise type.

Claims

1. An improved speech recognition system comprising:

a speech recognizer; and

a source normalization model coupled to said recognizer;

said model derived by a method of source normalization training for HMM modeling of speech comprising the steps of;

(a) providing an initial model;

(b) on said initial model or following new models performing the following steps to get a new model;

b₁) estimation of intermediate quantities;

b₂) performing re-estimation to determine initial state probability, transition probability, mixture component probability and environment probability;

b₃) deriving mean vector and bias vector;

b₄) solving jointly for mean vector and bias vector using linear equations and determining variances and transformation;

b₅) replacing old model parameters for the calculated ones; and

(c) determining after a new model is formed if it differs significantly from the previous model and if so repeating steps b₁ -b₅.

2. The method of claim 1 wherein in step b₁ estimation intermediate quantities is determined by α

_t (j,e)Δ

p(o₁^t,θ

_t =j,φ

=e|λ

),

β

_t (j,e)Δ

p(o_t+1^T |θ

_t =j,φ

=e,λ

), and γ

_t (j,k,e)Δ

p(θ

_t =j,ξ

_t =k,φ

=e|O,λ

).

3. The method of claim 2 wherein step b₂ the initial state probability is determined by ##EQU13## transition probability is determined by ##EQU14## mixture component probability is determined by ##EQU15## and environment probability is determined by ##EQU16##

4. The method of claim 2 wherein step b₃ deriving mean vector and bias vector is determined by

5. The method of claim 2 wherein step b₄ equations are used for solving jointly and equation ##EQU17## is used to determine variance and equations Z_je.sup.(m) =W_je.sup.(m) R_je (m), ##EQU18## are used to determine transformation.

6. A method of speech recognition comprising:

source normalization training for HMM modeling of speech comprising the steps of;

(a) providing an initial model;

(b) on said initial model or following new models performing the following steps to get a new model;

b₁) estimation of intermediate quantities;

b₂) performing re-estimation to determine initial state probability, transition probability, mixture component probability and environment probability;

b₃) deriving mean vector and bias vector;

b₄) solving jointly for mean vector and bias vector using linear equations and determining variances and transformation;

b₅) replacing old model parameters for the calculated ones; and

(c) determining after a new model is formed if it differs significantly from the previous model and if so repeating steps b₁ -b₅ ;

receiving an input signal; and

comparing said input signal to said new model.