Fractional excitation and other efficient transformed codebooks for CELP coding of speech

The authors deal with the design and efficient codebooks in terms of quality and complexity for code excited linear prediction (CELP) coding of speech. For this purpose, a transformed codebook approach is proposed and it is shown that several well-known CELP schemes can be formulated in this way. Comparative results are given for different transformation matrices, including a new method called fractional excitations. Another technique based on singular value decomposition (SVD) of the impulse response matrix is introduced. In this technique an innovation sequence appears as a linear combination of a set of eigenvectors with optimally vector quantized coefficients. Preliminary results are presented for this new SVD-CELP technique. The authors describe a multi-codebook approach which enabled a significant bit rate reduction of the speech coder submitted to GSM half rate standardization