A Front-End Processing Using Subspace-Based Speech Enhancement over Mobile Devices

Abstract

Abstract-The paradigm multi-stream has been shown to result in features combined that can help to increase the robustness of distributed speech recognition (DSR) in the mobile communications. In this paper, we employs a combination of post proceeded Mel-cepstral coefficients (MFCCs) and line spectral frequencies features (LSFs) projected in linear discriminate analysis (LDA) space. The experiments performed on the Aurora 2.0 database using multi-condition training set show that, even with fewer parameters, the proposed front-end provides comparable recognition results to the standard ETSI WI008 advanced front-end, nowadays available in the vocal commands of the GSM mobile communications, while achieving higher accuracy when the signal-to-noise ratio (SNR) is very low.Index Terms-Distributed speech recognition, linear discriminate analysis, multi-stream hidden Markov models, Mel-cepstral coefficients, line spectral frequencies.