Recently a number of researchers have reported a significant reduction in errors for speech recognition with different training and testing spectra (e.g. for change of microphone) using Relative Spectral (RASTA) approaches. In these studies, log spectral or cepstral coefficients are temporally filtered to reduce the effects of change in microphone, telephone channel, room acoustics, etc. These effects can be modeled as resulting from the convolution of some sequence with the speech data, resulting in a stationary or slowly-varying additive component in the log spectral domain. A similar approach has been used to reduce errors that are additive in the power spectral domain [3]. Unfortunately, in the more general case noise is both additive and convolutional; in particular, any real speech input includes both the effects of environmental echo response and microphone impulse response, as well as additive noise. Practical RASTA-based systems need to handle both effects simultaneously. In this paper, we report results from a series of new experiments in this problem domain. These results appear to show that a fairly simple refinement of the original RASTA approach provides some robustness to noise in addition to the robustness to convolutional effects.
Cite as: Morgan, N., Hermansky, H. (1992) RASTA extensions: robustness to additive and convolutional noise. Proc. ETRW on Speech Processing in Adverse Conditions, 115-118
@inproceedings{morgan92_spac,
author={Nelson Morgan and Hynek Hermansky},
title={{RASTA extensions: robustness to additive and convolutional noise}},
year=1992,
booktitle={Proc. ETRW on Speech Processing in Adverse Conditions},
pages={115--118}
}