Publication Details
Title: More Robust J-RASTA Processing Using Spectral Subtraction and Harmonic Sieving
Author: H. Ogawa
Group: ICSI Technical Reports
Date: August 1997
PDF: ftp://ftp.icsi.berkeley.edu/pub/techreports/1997/tr-97-031.pdf
Overview:
We investigated spectral subtraction (SS) and harmonic sieving (HS) techniques as preprocessing for J-RASTA processing to achieve more robust feature extraction for automatic speech recognition. We confirmed that spectral subtraction improved J-RASTA processing, and showed that harmonic sieving additively improved J-RASTA+SS. We investigated the performance with the Bellcore isolated digits task corrupted with car noise (additive noise) and linear distortion filter (convolutional noise). The J-RASTA+SS+HS system reduces the word error rate by 39% given pitch estimated from clean speech, and 35% given pitch estimated from corrupted speech. The system was also tested with several kind of noises from the NOISEX92 database; each noise sample was added with speech for a resulting of 0dB signal to noise ratio. SS significantly reduced word error rate for all type of noises (white noise 39%, pink noise 51%, car noise 78%, tank noise 59%, and machine gun noise 19%). Given correct pitch, HS additively reduced the word error rate for the first three noises (white noise 7%, pink noise 16%, and car noise 17%).
Bibliographic Information:
ICSI Technical Report TR-97-031
Bibliographic Reference:
H. Ogawa. More Robust J-RASTA Processing Using Spectral Subtraction and Harmonic Sieving. ICSI Technical Report TR-97-031, August 1997
Author: H. Ogawa
Group: ICSI Technical Reports
Date: August 1997
PDF: ftp://ftp.icsi.berkeley.edu/pub/techreports/1997/tr-97-031.pdf
Overview:
We investigated spectral subtraction (SS) and harmonic sieving (HS) techniques as preprocessing for J-RASTA processing to achieve more robust feature extraction for automatic speech recognition. We confirmed that spectral subtraction improved J-RASTA processing, and showed that harmonic sieving additively improved J-RASTA+SS. We investigated the performance with the Bellcore isolated digits task corrupted with car noise (additive noise) and linear distortion filter (convolutional noise). The J-RASTA+SS+HS system reduces the word error rate by 39% given pitch estimated from clean speech, and 35% given pitch estimated from corrupted speech. The system was also tested with several kind of noises from the NOISEX92 database; each noise sample was added with speech for a resulting of 0dB signal to noise ratio. SS significantly reduced word error rate for all type of noises (white noise 39%, pink noise 51%, car noise 78%, tank noise 59%, and machine gun noise 19%). Given correct pitch, HS additively reduced the word error rate for the first three noises (white noise 7%, pink noise 16%, and car noise 17%).
Bibliographic Information:
ICSI Technical Report TR-97-031
Bibliographic Reference:
H. Ogawa. More Robust J-RASTA Processing Using Spectral Subtraction and Harmonic Sieving. ICSI Technical Report TR-97-031, August 1997