| Presentation | 2004/12/14 Speech Recognition Adopting Compensated Acoustic Likelihood based on Noise Models. Shoei SATO, Kazuo ONOE, Akio KOBAYASHI, Toru IMAI, |
|---|---|
| PDF Download Page | PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | To improve recognition accuracy for speech uttered in a noisy environment, this paper proposes a new compensation method for acoustic scores in the - Viterbi search. In this method, to cope with wider varieties of background noise whose characteristics change rapidly, a confidence factor is obtained as a posterior probability of speech models or a likelihood ratio between speech models and noise models. This confidence factor represents the reliability of the acoustic score for the input speech. In decoding, weight of the acoustic score at a noisy frame is reduced according to the value of the confidence factor. An experiment with broadcast news transcription showed that this method reduced word errors for input speech with lower SNR values (0-5dB). The greatest reduction of word errors, by 20%, was obtained at an SNR of 0dB. This paper also proposes a modification of the compensation, which improved the recognition performance at a higher SNR of 10 dB. The proposed method is also applied to recognition of a noisy sports program. The results showed the method improved accuracy of keywords that is important for automatic meta-data extraction. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | speech recognition / noisy environment / acoustic score / compensation |
| Paper # | NLC2004-58,SP2004-98 |
| Date of Issue |
| Conference Information | |
| Committee | NLC |
|---|---|
| Conference Date | 2004/12/14(1days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | |
| Vice Chair | |
| Secretary | |
| Assistant | |
| Paper Information | |
| Registration To | Natural Language Understanding and Models of Communication (NLC) |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Speech Recognition Adopting Compensated Acoustic Likelihood based on Noise Models. |
| Sub Title (in English) | |
| Keyword(1) | speech recognition |
| Keyword(2) | noisy environment |
| Keyword(3) | acoustic score |
| Keyword(4) | compensation |
| 1st Author's Name | Shoei SATO |
| 1st Author's Affiliation | NHK Science and Technical Research Laboratories() |
| 2nd Author's Name | Kazuo ONOE |
| 2nd Author's Affiliation | NHK Science and Technical Research Laboratories |
| 3rd Author's Name | Akio KOBAYASHI |
| 3rd Author's Affiliation | NHK Science and Technical Research Laboratories |
| 4th Author's Name | Toru IMAI |
| 4th Author's Affiliation | NHK Science and Technical Research Laboratories |
| Date | 2004/12/14 |
| Paper # | NLC2004-58,SP2004-98 |
| Volume (vol) | vol.104 |
| Number (no) | 539 |
| Page | pp.pp.- |
| #Pages | 6 |
| Date of Issue | |