In our receutly reported work, we have observed some difference in recognition performance using our proposed method of feature computationn when comipared to features comiputed using the traditional mel-filterbank analysis. In the alternate method of feature computation, we use a spectral smoothing procedure which is very similar to weighted overlapped segment averaging (WOSA) method of spectral estimation. In this paper, we study the signal processing of the above mentioned feature computation methods, and point out to the differences between the two methods, and the effect of these differences on the recognition performance.
Cite as: Sinha, R., Umesh, S. (2003) A study into front-end signal processing for automatic speech recognition. Proc. Workshop on Spoken Language Processing, 87-93
@inproceedings{sinha03_wslp,
author={Rohit Sinha and S. Umesh},
title={{A study into front-end signal processing for automatic speech recognition}},
year=2003,
booktitle={Proc. Workshop on Spoken Language Processing},
pages={87--93}
}