A low delay code excited linear predictive speech coding algorithm at 8 kbits-per-sec and lower rate is considered. We try to improve the performance of speech coder with frame type dependent modification of the synthesis filter. We first classify frames into 3 groups: voiced, unvoiced and onset. On voiced and unvoiced frame, spectral envelope the synthesis filter is adapted to the phonetic characteristics. On transition frame from unvoiced to voiced, the synthesis filter is interpolated with the bias filter. The objective test shows improved SNRSEG and SPSEG values at 8 kbits-per-sec. The subjective test at 4.8 kbits-per-sec shows more preference of proposed coder than DoD-CELP coder. The Proposed coder produces better performance than the conventional CELP type coder with 85\% delay reduction.
