Intelligibility evaluation of a 4.8 kb/s residual-excited linear prediction coder
Mamoru Nakatsui, Paul G. Mermelstein, Dale C. Stevenson · The Journal of the Acoustical Society of America · 1980
Residual-Excited Linear Prediction (RELP) coding offers an alternative to Linear Prediction Coding in environments where background noise and transmission errors are likely to degrade the intelligibility of the received speech. The RELP evaluated here encodes the low-frequency component of the linear-prediction residual into two sub-bands. The component of the residual above 800 Hz is regenerated at the receiver, combined with the received low-frequency residual and used to drive the LPC synthesis filter. Sub-band coding of the low-frequency residual at 2 bits per sample yields quality comparable to coding the same residual with a single quantizer at 3 bits per sample. Evaluation with the aid of the Diagnostic Rhyme Test yields 95% intelligibility, significantly higher than most conventional linear-prediction coders. The intelligibility advantage derives from avoiding the voiced/unvoiced and pitch tracking decisions required for LPC and significantly increasing the time-resolution with which the excitation signal is encoded. Speech quality is midway between 3 and 4 bit log PCM coding. In 10-dB SNR background noise, intelligibility of the coded speech drops to 84% from 89% for the uncoded signal. The coder appears particularly suitable for use in noisy environments such as airplane to satellite communication where conventional LPC coders suffer serious degradations. [Work supported in part by the Communications Research Centre, Canada.