Investigation into a Mel subspace based front-end processing for robust speech recognition

This paper addresses the issue of noise reduction applied to robust large-vocabulary continuous-speech recognition (CSR). We investigate strategies based on the subspace filtering that has been proven very effective in the area of speech enhancement. We compare original hybrid techniques that combine the Karhonen-Loeve transform (KLT), multilayer perceptron (MLP) and genetic algorithms (GAs) in order to get less-variant Mel-frequency parameters. The advantages of these methods include that they do not require estimation of either noise or speech spectra. To evaluate the effectiveness of these methods, an extensive set of recognition experiments are carried out in a severe interfering car noise environment for a wide range of SNRs varying from 16 dB to -4 dB using a noisy version of the TIMIT database.