Phonetically-driven CELP coding using self-organizing maps

Results obtained by combining SOMs (self-organizing maps) with CELP (code excited linear prediction)-type speech coders are presented. The emphasis has been on exploiting the structured codebook that a SOM provides. Three different research directions have been addressed. The authors have tested, on a 3400-bit/s PS-VXC (phonetically segmented vector excitation coding), the use of information from SOMs as a flexible strategy for phonetic segmentation. The efficient encoding of the LPC (linear predictive coding) envelope was proposed to reduce the bit rate in a PS-VXC down to 3100 bit/s without perceptible degradations in the final quality. This efficient representation of LPC parameters was also combined with a phonetic-classification strategy to provide an improved 2400 bit/s CELP. Finally the authors present an efficient procedure for exploiting SOMs for a fast search quantization procedure that greatly reduces the complexity for VQ (vector quantization) of the spectral envelope.<>