Wavelet-based voiced/unvoiced classification algorithm

A new wavelet-based algorithm for classification of speech into voiced and unvoiced segments is presented. The algorithm is based on statistical analysis of the frequency distribution of the average energy in the wavelet domain, and on the short-time zero-crossing rate of the speech signal. First, the ratio of the average energy in the wavelet low-bands to that in the wavelet highest-band for each speech segment is computed using a 4-level dyadic wavelet transform, and compared to a predetermined threshold. This is followed by measuring the zero-crossing rate of the segment and comparing it to a threshold equal to the median of the zero-crossing rates. An experimentally verified criterion based on the above two comparison processes is then applied to obtain the voicing decision. The performance of the algorithm has been evaluated using a large speech database. The algorithm is shown to perform well in the cases of both clean and noise-degraded speech.