Title:
【発明の名称】音声認識用学習方式
Document Type and Number:
Japanese Patent JP2545961
Kind Code:
B2
Abstract:
PURPOSE:To facilitate learning with sound elements and syllables as units by iteratively correcting the load coefft. of a neutral net according to the error between the sound production pattern converted to sound element statement and the learning pattern subjected to the sound element statement by the neutral net. CONSTITUTION:The standard pattern of a voice pattern memory section 1 is converted to the pattern of the sound element statement in the neutral net section 4 and is supplied together with the learning pattern B subjected to the sound element statement of a sound element statement pattern memory section 2 to a DP matching section 5. The error dk of the pattern B from the matching section 5 and the converted pattern of the pattern Bk matched by the converted pattern is calculated by an error pattern calculating section 6. The load coefft. of the net section 4 is corrected by a neutral net correcting section 7 in accordance with this error dk. The similar correction is iterated until the result Dk of the calculation of an error relation calculating section 8 attains the prescribed value or below. As a result, the high-accuracy voice recognition device which allows automatic learning of the neutral net having the capacity to make identification with high accuracy from the voice pattern of the known sound production and allows the easy learning of the sound elements and syllables as units is obtd.
Inventors:
WATANABE TAKAO
Application Number:
JP184889A
Publication Date:
October 23, 1996
Filing Date:
January 06, 1989
Export Citation:
Assignee:
NIPPON ELECTRIC CO
International Classes:
G06F3/16; G06F15/18; G06N3/00; G10L15/06; G10L15/08; G10L15/16; G10L17/06; (IPC1-7): G10L9/10; G06F15/18; G10L3/00
Domestic Patent References:
| JP61102698A | ||||
| JP1291298A |
Other References:
【文献】日本音響学会講演論文集,平成元年3月,2-P-▲○27▼,P.283~284
Attorney, Agent or Firm:
Naoki Kyomoto (2 outside)