Trellis Code Excited Linear Prediction (TCELP) Speech Coding

This memo considers using extending techniques for sources with memory of the trellis-based scalar-vector quantizer (TB-SVQ) as efficient solutions to the embedding excitation codebook search problem in CELP speech coders. These solutions motivate a 24 kbit/s narrowband CELP coder, referred to as the low-delay trellis code excited linear prediction (LD-TCELP) coder, which utilizes its granular structure (from the underlying trellis code) and its codebook shaping capability (from the ideas of scalar-vector quantization and prediction/precoding) to effectively handle excitation vectors in the 24-dimensional space. This code has an algorithmic delay of 3 msec and, using such a novel codebook structure, it can capitalize on the higher dimensional vector quantization gains. In experiments we conducted, this LD-TCELP coder outperforms two toll-quality standardized coders: ITU-T G.728 16 kbit/s LD-CELP and ITU-T G.726 32 kbit/s ADPCM. The proposed LD-TCELP coder can also utilize a low-complexity codebook search approach. To the minimum extent, its computational complexity is about 75% of that of ITU-T G.728 LD-CELP.