Reduced complexity conditional entropy coding of LSP parameters

Line spectral pair (LSP) is a major representation scheme for LPC parameters of speech. In this paper, we propose a scheme that utilizes high-order correlations with a reduced complexity. VQ approaches also exploit high-order correlation, but it suffers from high sensitivity to different speech environments. This new scheme is based on scalar quantization plus entropy coding which can give more robust performance. In order to simplify the high-order conditional probability information, conditioning state reduction technique is provided. Due to the state reduction, 2nd order entropy coding is possible using only one minute of speech data with a significant performance gain. We achieve 1 dB log spectral distortion with about 24 bits/frame an average, where 27 bits/frame are needed for VQ at the similar distortion level.