Sixth International Conference on Spoken Language Processing (ICSLP 2000)

Beijing, China
October 16-20, 2000

Large Vocabulary Continuous Speech Recognition Under Real Environments Using Adaptive Sub-Band Spectral Subtraction

Masahiro Fujimoto, Jun Ogata, Yasuo Ariki

Department of Electronics and Informatics, Ryukoku University, Seta, Otsu-shi, Shiga, Japan

In this study, we propose an Adaptive Sub-Band Spectral Subtraction (ASBSS) method which can vary noise subtraction rate according to SNR in frequency bands at each frame. In the conventional Spectral Subtraction(SS), speech spectral is estimated by adjusting noise subtraction rate according to SNR. In general, SNR is defined and computed as the average over all the input speech signal. However, even if the noise is stationary, SNR varies according to speech energy. Therefore the subtraction rate of noise spectral should be adjusted according to the segmental SNR. This method is called Adaptive SS(ASS). Considering difference of spectral features such as vowel and consonant, the subtraction rate of noise spectral should be adjusted according to the sub-band SNR. This idea leads to the ASBSS method we propose in this paper. In order to evaluate the proposed method, we carried out Large Vocabulary Continuous Speech Recognition experiments and compared the results by our method with the conventional method in word accuracy.


Full Paper

Bibliographic reference.  Fujimoto, Masahiro / Ogata, Jun / Ariki, Yasuo (2000): "Large vocabulary continuous speech recognition under real environments using adaptive sub-band spectral subtraction", In ICSLP-2000, vol.1, 305-308.