Classified conditional entropy coding of LSP parameters

Summary form only given. A new LSP speech parameter compression scheme is proposed which uses conditional probability information through classification. For efficient compression of speech LSP parameter vectors it is essential that higher order correlations are exploited. The use of conditional probability information has been hindered by high complexity of the information. For example, a LSP vector has 34 bit representation at 4.8 K bps CELP coding (FS1016 standard). It is impractical to use the first order probability information directly since 2³⁴≈1.7×10¹⁰ number of probability tables would be required and training of such information would be practically impossible. In order to reduce the complexity, we reduce the input alphabet size by classifying the LSP vectors according to their phonetic relevance. In other words, speech LSP parameters are classified into groups representing loosely defined various phonemes. The number of phoneme groups used was 32 considering the ambiguity of similar phonemes and background noises. Then conditional probability tables are constructed for each class by training. In order to further reduce the complexity, split-VQ has been employed. The classification is achieved through vector quantization with a mean squared distortion measure in the LSP domain