8 kb/s low delay CELP with feedback vector quantization

Abstract

A 8 kb/s speech coder in which the one-way coding delay is less than 3.75 ms is described. The proposed coder is a LDCELP (low delay code excited linear predictive coders) incorporating a FSVQ (finite state vector quantizer). A FSVQ is a switched vector quantizer with a backward selecting mechanism, and it is able to improve coding performance without increasing delay. Experimental results confirm that the incorporation of a 4-state FSVQ with a 1-frame state transition memory into a LDCELP provides 0.6 dB improvement in signal-to-noise ratio, and that the incorporation of a 4-state FSVQ with 3-frame state transition memories into a LDCELP provides 1.0 dB improvement. The signal-to-noise ratios of these two proposed LDCELPs are 13.7 dB and 14.1 dB, respectively.<>