講演抄録/キーワード |
講演名 |
2008-03-20 15:15
[ポスター講演]Initial Evaluation of the Drivers' Japanese Speech Corpus in a Car Environment ○Kousuke Hiraki・Takahiro Shinozaki・Koji Iwano・Agnieszka Betkowska・Koichi Shinoda・Sadaoki Furui(Tokyo Inst. of Tech.) SP2007-202 |
抄録 |
(和) |
Car navigation systems are getting more and more popular and many of them equip a speech recognition system as hands-free interface. However, the speech input interface is not widely used because of insufficient recognition performance. In order to improve the recognition performance and make the speech interface more practical, a real-car-environment speech corpus "Drivers' Japanese Speech Corpus in a Car Environment" is under construction by a project led by the Japan Ministry of Economy, Trade and Industry. In this study, we used the command task portion of the corpus recorded under three conditions: while idling, running in a city, and running on a highway. We used the data from the corpus only as a test set and made a recognition system by optimally combining several existing corpora with several noise robustness techniques. Experimental results show that using a HMM trained on multiple conditions with spectral subtraction is the best for the car noises. Recognition performance was largely improved and more than 90% word accuracy was achieved for all the recording conditions. In particular, over a 50% absolute improvement in accuracy was observed for speeches given by female speakers uttered on a highway. |
(英) |
Car navigation systems are getting more and more popular and many of them equip a speech recognition system as hands-free interface. However, the speech input interface is not widely used because of insufficient recognition performance. In order to improve the recognition performance and make the speech interface more practical, a real-car-environment speech corpus "Drivers' Japanese Speech Corpus in a Car Environment" is under construction by a project led by the Japan Ministry of Economy, Trade and Industry. In this study, we used the command task portion of the corpus recorded under three conditions: while idling, running in a city, and running on a highway. We used the data from the corpus only as a test set and made a recognition system by optimally combining several existing corpora with several noise robustness techniques. Experimental results show that using a HMM trained on multiple conditions with spectral subtraction is the best for the car noises. Recognition performance was largely improved and more than 90% word accuracy was achieved for all the recording conditions. In particular, over a 50% absolute improvement in accuracy was observed for speeches given by female speakers uttered on a highway. |
キーワード |
(和) |
car navigation system / noise robustness / spectral subtraction / tree-structured clustering / / / / |
(英) |
car navigation system / noise robustness / spectral subtraction / tree-structured clustering / / / / |
文献情報 |
信学技報, vol. 107, no. 551, SP2007-202, pp. 93-98, 2008年3月. |
資料番号 |
SP2007-202 |
発行日 |
2008-03-13 (SP) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
SP2007-202 |
|