An efficient of Neural Address Predictor applies to Address Vector Quantisation codebook in speech processing

Generally characteristic of speech waveform is the continuous signal, which contains of voiced and unvoiced signal. Historically, speech waveform is coded by dividing it into frames; it is typically divided into 30 ms frame length, where each frame is coded separately. Speech is however created by a physical system and is substantially shaped by the vocal tract. As it is physically impossible for the vocal tract to move instantaneously from any state to any given state, trends should exist between successive vocal tract positions. In the coding techniques used in this paper, the vocal tract positions manifest themselves as Vector Quantised LSP coefficients. Although speech coding is an entity in its own right, strong links exist between image compression and speech compression. In this work, the Address-VQ technique which used in the image compression arena, have been applied to the compression of speech coded parameters. Furthermore the technique, called Neural Address Prediction, which is a lossy technique, also applied to encourage further reduce the bit rate. This work exploits the repetitiveness of the attribute of a single speaker to further reduce the bit rate. Preliminary results indicate that approximately more than 33% additional compression is achievable using Neural Address Prediction with Address Vector Quantisation codebook. As Neural Address Prediction is a lossy compression scheme, the error of prediction directly affects to the quality of synthesis speech especially in the voice frames.