Recently, there have been considerable researches that achieve high quality speech of vocoder at the rate of 2.4 to 4kbps. There are, however, some problems to reduce the bit rate to 4kbps or below while maintaining high speech quality. The reduction of bit rate causes the rapid degradation of speech quality in CELP(Code Excited Linear Prediction) vocoder and requires expensive computational complexity in MBE(Multi-Band Excitation) vocoder.
This paper proposes the 2.9kbps LP-SMBE(Linear Prediction - Simplified Multi-Band Excitation) vocoder that produces high speech quality with low computational complexity. This vocoder uses a new speech model, referred to as the LP-SMBE speech model, that represents spectral envelope with LPC spectrum and excitation spectrum with MBE speech model. This paper also suggests an efficient estimation method for pitch period and voiced/unvoiced decisions using normalized spectrum matching.
The performance of the proposed vocoder was compared with other vocoders on the aspects of speech quality and computational complexity. The preference test of speech quality has shown that LP-SMBE exhibits better speech quality than that of 4.8kbps DoD CELP. It has been shown that the vocoder has lower computational complexity than that of 4.8kbps DoD CELP and 2.4kbps IMBE vocoder.
