Speech vocoder presented in this paper classifies input speech into silent

unvoiced

voiced and mixed classes.Adaptive multi band classification algorithm is applied for unvoiced/voiced decision and active/inactive decision.Such algorithm has improved robustness

accuracy and flexibility of classification

and the quality of mixed speech is also maintained without coding of U/V decision vector.Spectra of unvoiced and voiced classes are encoded with different LSP quantization tables

respectively

and then a scale quantizer could replace a vector quantizer

therefore less complexity is achieved.Bit rates of the vocoder are maximum 2.4kbps and minimum 0.1kbps

and average bit rate is 1.4kbps.A real time software implementation of the vocoder has 0.3 second of system delay.A 40MHz TMS320C50 fixed point DSP is chosen for real time implementation of the decoder and synthesizer parts

and the average computation is 11.3MIPS.