Phonetic classification for a novel Mandarin vocoder

The algorithm presented in this paper is designed for a novel phonetic classified low bit rate Mandarin vocoder with natural speech quality and DSP implementation. Original speech is classified into four classes of voice-silence, unvoiced, voiced, and mixed. Bit rates allocated for silence class and unvoiced class can be reduced greatly without significant degradation in speech quality. The vocoder architecture can be simplified in bit-rates of maximum 2.4 kbps and average 1.4 kbps. This paper emphasizes on the phonetic classification algorithm of the vocoder which is the key to improve synthesized speech quality. Compared to several phonetic classification algorithms, the algorithm in this paper appears to be more effective, more accurate and less complex. With the synthesized speech quality being maintained, the complexity of whole vocoder is simplified and the IC implementation and DSP implementation can benefit.