Address-vector quantisation applied to speech coding

The use of address vector quantisation (VQ) in the compression of linear predictive coded (LPC) and line spectral pairs (LSP) speech parameters in a speaker dependent system are examined. Four speakers are investigated; two male and two female. The speech waveform is coded to LPC and LSP parameters using LPC techniques and is vector quantised using an unsupervised neural network, a Kohonen self organising feature map (KSOFM), to create a codebook of 128 entries. Address VQ is applied to the codebook and the data examined for recurring sequences to exploit redundancy. Preliminary results indicate that approximately 46% additional compression is achievable using this method. As Address VQ is a loss-less compression scheme, this reduction is achieved without any further reduction in speech quality.