Spectral subtraction in the time-frequency domain using wavelet packets

This paper deals with the problem of enhancement of speech signals in slowly varying background noise. The method suggested uses the wavelet packet transform together with a simple statistical model of the noise. As in spectral subtraction, information about the background is collected during non-speech intervals. The signal estimate is obtained with spectral subtraction in the time-frequency domain and a rule for choosing the most favorable time-frequency decomposition is suggested. The method is demonstrated with a real-data example.