Speech recognition by neural network adapted to reference pattern learning

• US 5,600,753 A
• Filed: 07/05/1994
• Issued: 02/04/1997
• Est. Priority Date: 04/24/1991
• Status: Expired due to Fees

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Iso, Ken-ichi
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): CHOWDHURY, INDRINAL
• Application Number:
  US08/270,416
• Time in Patent Office: 945 Days
• Field of Search: 381/41-45, 395/2, 395/2.41, 395/2.11, 395/2.68, 382/15
• US Class Current: 704/200
• CPC Class Codes: G06F 18/295 Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G10L 15/16 using artificial neural networks

Abstract

A speech recognition method according to the present invention uses distances calculated through a variance weighting process using covariance matrixes as the local distances (prediction residuals) between the feature vectors of input syllables/sound elements and predicted vectors formed by different statuses of reference neural prediction models (NPM'"'"'s) using finite status transition networks. The category to minimize the accumulated value of these local distances along the status transitions of all the prediction models is figured out by dynamic programming, and used as the recognition output. Learning of the reference prediction models used in this recognition method is accomplished by repeating said distance calculating process and the process to correct the parameters of the different statuses and the covariance matrixes of said prediction models in the direction of reducing the distance between the learning patterns whose category is known and the prediction models of the same category as this known category, and what have satisfied prescribed conditions of convergence through these calculating and correcting processes are determined as reference pattern models.

Claims

1. A pattern recognition method for recognizing syllables and sound elements on the basis of the comparison of input time series patterns expressed as feature vectors of the syllables and sound elements with reference pattern models using a finite status transition network, in which each status of said finite status transition network has a predictor, comprising the steps of:

(a) calculating, in each predictor, a predicted feature vector at time t from a plurality of input feature vectors between time (t-1) and time (t-τ

_F) and a plurality of input feature vectors between time (t+1) and time (t+τ

_B), wherein said τ

_B and τ

_F are predetermined natural number;

(b) determining a local distance at every t between said input feature vectors and t-th status of said finite transition network by using said input feature vectors, said predicted feature vector and a covariance matrix which accompanies t-th status of said finite status transition network;

(c) calculating an accumulated value of said local distances for every reference pattern defined by said status of said finite state transition network;

(d) detecting a minimum of said accumulated values for every reference pattern; and

(e) outputting a category of the reference pattern corresponding to said minimum as a recognition result.

2. A speech recognition method, as claimed in claim 1, wherein initial values are set for the parameters of said predictor and said covariance matrix accompanying each status of said finite status transition network, said local distance between said input time series pattern, category of which is known, and said reference pattern model corresponding to the same category as said known category is calculated;

the parameters of said predictor and said covariance matrix of each state are iteratively corrected by using a gradient descent method; and

said reference pattern model said local distance for which satisfies predetermined conditions of convergence is thereby obtained.