A 2.4 kbit/s CELP speech codec with class-dependent structure

A 2.4-kbit/s analysis-by-synthesis speech codec based on a CELP (code-excited linear prediction) structure is presented. Several bit-rate reduction techniques, including the addition of a frame classifier and a frame-dependent codec structure, are used to improve the resulting speech quality. For voiced frames, the encoding is based entirely on the past excitation retrieved from an adaptive codebook using multitap gains with as many as seven taps. Unvoiced frames are encoded using only stochastic excitation. Finally, transition frames use both the adaptive and the stochastic codebooks. The results of subjective quality evaluation show that the open-loop classifier performs slightly better than the open-loop classifier, especially for female speakers. The best 2.4-kbit/s system uses the closed-loop classifier and a five-tap adaptive codebook for encoding voiced segments. This system achieve better subjective ratings than LPC-10e; however, it is still inferior to the Federal standard operating at double the bit rate, although for many speakers the difference was small.<>