Spectral peaks enhancement for extracting robust speech features

It is generally believed that the external noise added to speech signal corrupts speech spectrum and so speech features. This feature corruption degrades speech recognition systems performance. One solution to cope with the speech feature corruption is reducing the noise effects on the speech spectrum. In this paper, we propose to filter speech spectrum in order to enhance its spectral peaks in presence of noise. Then, we extract robust features from the spectrum with enhanced peaks. In addition, we apply the proposed filtering to another form of speech spectral representation known as modified group delay function (GDF). Phoneme and word recognition results show that MFCC features extracted from the spectrum with enhanced peaks are more robust to noise than MFCC derived from main noisy spectrum. In addition, MFCC features extracted from filtered GDF are more robust to noise than other MFCC features, especially in low SNR values.