Bit Rate Reduction of FS-1015 Speech Coder Using Fuzzy ARTMAP and KSOFM Neural Networks

The speech spectrum is very sensitive to linear predictive coding (LPC) parameters, so small quantization errors may cause unstable synthesis filter. Line spectral pairs (LSPs) are more efficient representations than LPC parameters. On the other hand, artificial neural networks (ANNs) have been used successfully to improving the quality and also reduction the computational complexity of speech coders. This work proposes an efficient technique to reduce the bit rate of FS-1015 speech coder, while improving the performance. In this way, LSP parameters are used instead of the LPC parameters. In addition, neural vector quantizers based on Kohonen self-organizing feature map (KSOFM), with a modified-supervised training algorithm, and fuzzy ARTMAP are also employed to reduce the bit rate. By using the mentioned neural vector quantizer models, the quality of synthesized speech, in terms of mean opinion score (MOS), is improved 0.13 and 0.26, respectively. The execution time of proposed models, as compared to FS-1015 standard, is also reduced 27% and 43%, respectively.