Method for production of speech reference templates

• US 4,590,605 A
• Filed: 12/14/1982
• Issued: 05/20/1986
• Est. Priority Date: 12/18/1981
• Status: Expired due to Term

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Hataoka, Nobuo, Ichikawa, Akira
• Primary Examiner(s): Kemeny, E. S. Matt
• Application Number:
  US06/449,660
• Time in Patent Office: 1,253 Days
• Field of Search: 381/41-50
• US Class Current: 704/245
• CPC Class Codes: G10L 15/063 Training

Abstract

In this speech recognition system, a set of templates for each phoneme includes clusters of speech patterns based on two speech features: "physical" features (formant spectra of men versus women) and "utterance" features (unvoiced vowels and nasalization), derived from a plurality of reference speakers.

Claims

1. A method for production of speech reference templates for phoneme groups for use in speech recognition, each of said phoneme groups containing at least one phoneme, the method comprising the steps of:

preparing first speech patterns with respect to speeches of plural speakers, said first speech patterns corresponding to a word which is stable by utterance;

classifying said speeches of said plural speakers into speech groups by use of differences in said first speech patterns caused by physical features of said speakers;

preparing second speech patterns with respect to said speeches of said plural speakers, said speech patterns corresponding to a phoneme group containing at least one phoneme;

subclassifying the speeches of said plural speakers in each of said speech groups into speech subgroups by the use of differences in said second speech patterns caused by features of speech occurring by utterance; and

selecting a plurality of second speech patterns among the second speech patterns in each of said speech subgroups and assigning said selected plurality of second speech patterns to be the speech reference template for each of said phoneme groups.

2. A method according to claim 1, wherein said classifying and subclassifying of the speeches of said plural speakers are carried out through hierarchical clustering in which, in each of said speech groups and speech subgroups, the speeches of said plural speakers are hierarchically grouped into a predetermined number of sets in terms of the degree of similarity in their speech patterns.

3. A method according to claim 1, wherein the predetermined number of sets in each of said speech groups and speech subgroups is decided on the basis of the ratio between a value related to the similarity among the speech patterns in each of said speech groups or speech subgroups and a value related to the similarity of the speech patterns among said speech groups or speech subgroups.