Adaptive pitch period decimation and its application in speech compression

This paper presents a new method of speech coding that takes advantage of the repetitiveness inherent in voiced speech. Voiced speech is broken into pitch period lengths (wavelets) and these signals are compared with one another to determine If two wavelets differ significantly. If the wavelets are significantly different, then they are encoded and transmitted; otherwise, the current wavelet Is not transmitted, and the next pitch period wavelet is compared. This results in encoding only a representative fraction of the speech signal and significantly lowers the number of bits required to transmit the signal. Pitch period determination is done by using the autocorrelation method and a median smoothing filter. The pitch period wavelets are preprocessed using a time weighted averaging method that allows concatenation of wavelets without sharp transitions at pitch boundaries, therefore reducing high frequency noise. Wavelets are compared using the Itakura distance measure, which is usually employed in speech recognition applications. The transmitted wavelets are encoded using a differential PCM method to further reduce the bit rate of the transmission. Unvoiced speech is encoded using an LPC method on a frame by frame basis. This results in high quality speech transmission at bit rates of approximately 3.8 kb/s.