Robust speech recognition based on multi-band spectral subtraction

Abstract

In order to reduce the degradation of the speech recognition accuracy while the testing condition are mismatched with the training condition around noisy environment, a kind of multi-band spectral subtraction has been proposed. The estimated noise signals were extracted from the first few frames of the noisy speech. The noisy speech and estimation of noise signals by the frequency were divided into non-overlapping M frequency bands. According to the SNR (signal-to-noise ratio) of noise speech in each frequency band, the band noise spectral subtraction parameters can be determined. The front-end speech enhancement module and the speech recognizer constitute a robust speech recognition system. The results of simulation experiments indicate that the recognition accuracy of multi-band spectral subtraction robust speech recognition system is obviously superior to the basic spectral subtraction in different signal-to-noise ratios and different noise's types.