Study on low bit and variable rate speech coding based on local cosine transform
LI Yong-li · Journal of China Institute of Communications · 2005
A 1.6kbit/s low bit and variable rate speech coder based on local cosine transform (LCT) algorithm for two-way conversational speech was designed for the first time. The VAD (voice activity detector) based on SVM (support vector machine) and the classification method of speech modes of the GSM half rate standard for active speech were adopted in the design of the variable rate coder. A little modification for speech mode was made where the moderately voiced mode and the strongly voiced mode were combined as a speech mode. The new combined speech mode was named as moderately and strongly voiced mode. A few segment vector quantizers of the local cosine transform coefficients for each speech mode and silence speech frame (background noise) were employed, and LGB algorithm was applied to design the codebooks. A tree fast search technique was used to select the vector of local cosine transform coefficients for each segment. The evaluation, using subject informal listening tests, indicated that the quality of the synthesized speech of the designed low bit and variable rate speech coder was comparative with that of the U. S. Federal Standard MELP vocoder (2.4 kbit/s). The subjective MOS mean opinion score of the designed speech coder was about 3.15. The new coder has higher robust than the U. S. Federal Standard MELP vocoder, which is suitable for speech coding in any environment.