Design of a variable half rate speech codec

Abstract

In this paper, we combined generalized analysis-by-synthesis (AbS) structure and an algebraic excitation scheme to propose a variable half rate speech codec. The proposed codec operates on the speech frame of 20 ms at a sampling rate of 8 kHz. Also we introduce an efficient voicing detection algorithm using both frame energy and acoustic parameters to detect the voiced, unvoiced and silence region. 4 kb/s speech codec for the voiced region uses the generalized AbS structure which is efficient for the low bit rate speech codec. The generalized AbS structure is used for the relaxed CELP (RCELP) and it can reduce the bitrate for the pitch information by linear interpolation of the pitch delay. LP coefficients are converted to line spectrum pairs (LSP) and quantized using a predictive 2-stage VQ. A low complexity algebraic codebook which uses a shifting method is used for the fixed codebook excitation. 4 kb/s speech codec for the unvoiced region is same as that for the voiced region, except that the latter uses an open loop pitch lag in the whole frame. Also 1 kb/s codec is designed for the silence and background noise region. Experiments show that the variable half rate codec provides the average bit rate of about 2.63 kb/s for the test sentences whose voice activity factor (VAF) is 0.47. The mean opinion score (MOS) of the proposed variable half rate codec is 3.69.