Title:
VOICE RECOGNITION DEVICE AND METHOD
Document Type and Number:
Japanese Patent JP3166708
Kind Code:
B2
Abstract:
PROBLEM TO BE SOLVED: To select an optimum speaker cluster and an HMM(hidden Markov model) or probability distribution showing the cluster for every recognition unit or every state of the HMM for speaker adaptation to a new speaker.
SOLUTION: A device for performing a voice recognition based on a hidden Markov model comprises a speaker tree memory 7 for storing a plurality of speaker trees in every HMM of recognition unit, a known node selecting means 4 for selecting an optimum node of the speaker trees by using the sound by a new speaker, a speaker frequency calculating means 5 for counting the frequency information of the terminal node of a partial tree having the selected node as parent node, an known node selecting means 6 for selecting an optimum node of the speaker trees with respect to the speaker trees of recognition unit not included in the adapted sound of the new speaker by use of the frequency information of terminal node, and a standard pattern updating means 8 for collecting the HMM corresponding to the nodes selected by the known node selecting means and the unknown node selecting means and storing it in a standard pattern memory part 3 as a standard pattern corresponding to the new speaker.
Inventors:
Kenichi Iso
Application Number:
JP15702398A
Publication Date:
May 14, 2001
Filing Date:
June 05, 1998
Export Citation:
Assignee:
NEC
International Classes:
G10L15/10; G10L15/06; G10L15/07; G10L15/14; G10L17/00; (IPC1-7): G10L15/10
Other References:
【文献】Proceedings of 1994 IEEE International Conference on Acoustics,Speech and Signal Processing,Vol.1,T.Kosaka et al,”Tree-Structured Speaker Clustering for Fast Speaker Adaptation”,p.I-245~I-248
Attorney, Agent or Firm:
Asamichi Kato