A study of low bit-rate linear prediction based speech coders for Indian dialects

Abstract

Speech coders based on the linear prediction model are widely in use today. This paper describes the algorithms of low bit-rate vocoders, viz. Code-Excited Linear Prediction (CELP) and Mixed Excitation Linear Prediction (MELP) and their performance for Indian dialects. A Linux platform has been used for execution of the vocoders. Mean Opinion Score testing has been performed with speech samples of Indian dialects and Indian-accented English. Waveform analysis has been done using Praat software. Results show that CELP which operates at a higher bit rate of 4.8 kb/s gives better quality output. MELP speech output is intelligible and suited for very low bit-rate (2.4 kb/s) applications. The results also highlight the need for better codebook tuning in Indian dialects.