Technical Program

Paper Detail

Paper:	SLP-P19.2
Session:	Model-based Robust Speech Recognition
Time:	Friday, May 19, 10:00 - 12:00
Presentation:	Poster
Topic:	Speech and Spoken Language Processing: Model-based robust Speech Recognition
Title:	Novel Model Compensation For Features Based on SNR-dependent Non-uniform Spectral Compression
Authors:	Geng-Xin Ning, South China University of Technology, China; Shu-Hung Leung, Kam-Keung Chu, City University of Hong Kong, China; Gang Wei, South China University of Technology, China
Abstract:	This paper proposes a novel model compensation method for a robust feature extraction technique based on SNR-dependent Non-uniform Spectral Compression (SNSC). The SNSC method is a spectral transformation which resembles human's intensity-to-loudness conversion and de-emphasizes the contributions from noisy spectral components to features. In this paper, we propose a new compressed mismatch function which models the effect of the noise onto the clean speech in the Log-spectral domain together with the SNSC. Based on this mismatch function, a new model compensation procedure is derived. The procedure needs a compensated model of no compression to start with. It is shown that the new model compensation using the Vector Taylor Series method (VTS) for the compensated uncompressed model, remarkable recognition performances at low signal-to-noise ratio (SNR) can be obtained for different additive noises at the expense of slight increase in the computational complexity in comparison with the VTS.