Fast LSP calculation and quantization with application to the CELP FS1016 speech coder

Line Spectrum Pair (LSP) representation is used for spectral quantization in the CELP FS1016 speech coder, where the LSPs are first calculated, and then quantized using 34-bit non-uniform scalar quantization. In the algorithm proposed in this paper, computational complexity is decreased by searching the zero-crossings on the grid formed by the values of the quantization tables. As the actual LSPs are not calculated, two criteria to select the "closest" quantized LSPs are proposed. These criteria take into account the interaction between successive LSPs. The efficiency and reliability of the proposed algorithm are improved using the interlacing property of the LSPs and knowledge of the direction of the sign-change at every zero-crossing. The proposed algorithm is compared with the existing Kabal's algorithm (followed by quantization), showing similar quantization performance. The computational complexity on a fixed-point DSP56001 implementation is reduced by 66 %, using the proposed algorithm.