Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
SPEECH RECOGNITION METHOD AND MODEL DESIGN METHOD THEREFOR
Document Type and Number:
Japanese Patent JPH10282986
Kind Code:
A
Abstract:

To realize an adaptive speech recognition system by speaker classification and to easily obtain high performance speech recognition by using fuzzy inference, then estimating the output probability of a hidden Markov model(HMM) from information of similarities of respective classifications and the output probability in respective classifications.

First of all, speech samples to be learned are collected (S 402). Then, the classifications of the speech samples are decided (A 403). The classifications of the speech sample are decided heuristically, or the optimum classifications are decided from the characteristics of the speech data 420 to be learned. Then, the speech data are answered to respective classifications according to the decided classifications to be divided (S 405). Then, HMM parameters are learned (S 411-419) by using the speech data 421-429 at every classification. At this time, values of preliminary learning are used as state transition probabilities of respective HMMs. Then, adjustment for constituting a whole system is performed from the HMMs learned at every classification (S 406). Methods using the fuzzy theory and the method using a neural network, etc., are exemplified therefor.


Inventors:
NAKAGAWA TOMOHITO
MAEJIMA HIDEO
Application Number:
JP1997000086486
Publication Date:
October 23, 1998
Filing Date:
April 04, 1997
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
HITACHI LTD
International Classes:
G10L15/10; G10L15/02; G10L15/08; G10L15/14; (IPC1-7): G10L3/00; G10L3/00; G10L9/10
Attorney, Agent or Firm:
磯村 雅俊



 
Previous Patent: HEARING SUPPORT DEVICE

Next Patent: SPEECH RECOGNITION DEVICE