Noise robust speech recognition by combining speech enhancement in the wavelet domain and Lin-log RASTA

For improving noise robustness of speech recognition under adverse noise environment, a method of noise robust speech recognition, which combines discrete wavelet transform (DWT), wavelet packet decomposition (WPD) and Lin-log RASTA, is researched in this paper. After one scale of DWT was employed for noisy speech, this method used three scales of DWT and three scales of WPD for the low frequency signal and the high frequency signal, respectively. Multithresholds processing and decision of unvoiced sounds and voiced sounds were also adopted in order to improve the performance of denoising. The Lin-log RASTA coefficients were then computed from the enhanced speech as feature vectors. Cepstral mean subtraction (CMS) was used for compensating the speech distortion and residual noise of the above processing. Experimental results indicate that this method performs better for digital speech recognition than Lin-log RASTA, Spectral Subtraction+ Lin-log RASTA, mel-frequency cepstral coefficients (MFCC) and RASTA.