Title:
NOISE RESISTING PHONEME MODEL GENERATING SYSTEM
Document Type and Number:
Japanese Patent JP3247746
Kind Code:
B2
Abstract:
PURPOSE: To improve voice recognition performance by adding noise HMM and phoneme HMM, which is generated based on the voice information beforehand collected under a noise free condition, in a spectrum region and generating a phoneme model suitable to uttered voice environment.
CONSTITUTION: In a cepstrum region, a phoneme HMM(Hidden Markov Model) in a noise free condition and a noise HMM in a noisy condition are beforehand generated, are respectively cosine transformed and respective logarithmic spectra are computed. Then, they are exponential transformed, respective linear spectra are computed and output probabilities of phoneme and noise are obtained by convolution. The noise convoluted phoneme HMM is exponential transformed, is inverse cosine transformed and a phoneme HMM is generated. Since the noise model of the uttered voice environment is used, a good phoneme HMM, which is suitable to the uttered voice environment, is generated and a high phoneme recognition rate is obtained.
Inventors:
Kiyohiro Shikano
Yasuhiro Minami
Tatsuo Matsuoka
Frank martin
Yasuhiro Minami
Tatsuo Matsuoka
Frank martin
Application Number:
JP568893A
Publication Date:
January 21, 2002
Filing Date:
January 18, 1993
Export Citation:
Assignee:
Nippon Telegraph and Telephone Corporation
International Classes:
G10L15/06; G10L15/065; G10L15/14; G10L15/20; (IPC1-7): G10L15/14; G10L15/06; G10L15/20
Domestic Patent References:
| JP283596A |
Other References:
【文献】Proc IEEE Int Conf Acoustics、Speech、and Signal Process(ICASSP)1992、Vol.1 I-489~I-492頁「ADAPTATION OF THE HMM DISTRIBUTIOS:APPLICATION TO A VQ CODEBOOK AND TO A NOISY ENVIRONMENT」AU:FRANGOULIS E D、GAGANELIS D A
Attorney, Agent or Firm:
Takashi Sawai
Previous Patent: ELECTROSTATIC IMAGE DEVELOPING TONER
Next Patent: COMMUNICATION METHOD FOR ISDN DATA TERMINAL EQUIPMENT
Next Patent: COMMUNICATION METHOD FOR ISDN DATA TERMINAL EQUIPMENT