ISCA Archive ICSLP 1992
ISCA Archive ICSLP 1992

A mixed Gaussian-stochastic code book for CELP coder in LSP speech coding

Najib Naja, Jean Marc Boucher, Samir Saoudi

In a last few years, the Code Excited Linear Predictive (CELP) coder has demonstrated that such coders offer considerable promise for producing high quality synthetic speech at bit rates between 4.8 and 16 kbit/s. Tremendous efforts have been made to improve the CELP coder performance or to reduce its implementation cost. The CELP performance has been studied for different kind of code books such as Gaussian code books, algebraic code books, ternary code books ... In this paper, the Gaussian and a mixed Gaussian-stochastic code books were compared at different sizes, while the Line Spectrum Pairs (LSP) were used instead of Log Area Ratios (LAR) which provide efficient representation of the synthesis filter used in the CELP. This new stochasic code book was designed for various sizes, using the K-means algorithm applied to a large training data. Its first half is a Gaussian code book and the other half is obtained from the prediction error (after standardizing to unit variance) of speech signals when the skewness pi and the kurtosis 02 parameters do not correspond to the Gaussian law. For the same size, the new stochastic code book offers a better performance than the Gaussian one, about 0.5 dB for the same number of bits.