Speech recognition with wavelet spectral subtraction in real noisy environment

In this paper we focused the effectiveness of the wavelet spectral subtraction in noisy speech recognition. For this purpose, Fourier spectral subtraction is a conventional effective technique, for example. It is a suitable technique for stationary noise reduction (ex. white Gaussian like noise), because the short-time Fourier transform provides a uniform time-frequency resolution on each frequency band. However, it cannot reduce sudden noise effectively. On the other hand, the wavelet transform may be a suitable technique for sudden signal analysis, etc. (non-stationary signal analysis), because it admits a non-uniform time-frequency resolution on each frequency band. Therefore, we report to provide effective performance of noise reduction using the Fourier spectral subtraction the wavelet spectral subtraction and the microphone array steering in real noisy environments on EUROSPEECH2003. However, it was not clear that what kind of noise characteristics could be reduced with the wavelet spectral subtraction. In this paper, to cope with this problem, we evaluated the performance of the wavelet spectral subtraction and the Fourier spectral subtraction in various noisy environments. As a result of evaluation experiments, we confirmed that the wavelet spectral subtraction could effectively reduce sudden noise or higher frequency noise than the Fourier spectral subtraction.