Smoothed spectral subtraction for a frequency-weighted HMM in noisy speech recognition

The paper proposes improved methods of smoothed spectral subtraction to enhance the recognition performance of a frequency weighted HMM (HMM-FW) in very noisy environments. The conventional spectral subtraction tends to produce discontinuity in estimated power spectra. This distortion is undesirable for HMM-FW which uses group delay spectra as feature vectors. In order to remove this distortion, the paper proposes two frequency smoothing methods in log spectral domain: (1) a low pass filtering by DCT; and (2) a weighted minimum mean square error method (WMSE) which fits cosine series to an estimated log power spectrum. The results show that the smoothers are very effective under very noisy conditions, especially for the frequency weighted HMM. The WMSE method combined with HMM-FW achieves the highest recognition accuracies, for instance, improving recognition rate from 68% to 88% at -6 dB SNR of car noise