Improved 4.8 kb/s CELP coding using two-stage vector quantization with multiple candidates (LCELP)

The authors propose 4.8-kb/s LCELP (learned code excited LPC) coding. In order to improve conventional CELP speech quality with relatively small computation amount at around 4.8 kb/s. in 4.8-kb/s LCELP, the excitation signal is quantized by a two-stage vector quantizer (VQ) with multiple candidates. The conventional two-stage VQ can significantly reduce the computational amount for the excitation codebook search. However, the conventional two-stage VQ performance is usually lower than that for a single-stage VQ with the same number of bits. In order to overcome this problem and to produce high quality speech, the authors use a two-stage VQ with multiple candidates. The optimum combination of first and second code vectors is searched for among multiple candidates. Two excitation codebooks are designed iteratively, using a speech database. An experimental result shows that the 4.8-kb/s LCELP achieves an average of 11.8 dB segmental SNR, which is 1.6-dB higher than that for a 4.8 kb/s conventional CELP. Informal listening tests indicate that the 4.8 kb/s LCELP speech quality is significantly higher than that for the 4.8 kb/s conventional CELP coder