A 2.4/1.2kb/s speech coder with noise pre-processor

Speech coding at very low bit rates has many applications such as answering machines, IP telephony, mobile communications, military communications etc. In this paper we describe a speech coder capable of operating at both 2.4 and 1.2kb/s, and produces good quality synthesised speech. The basic principle of the coder is based on frequency domain vocoding method where the LP excitation is classified into voiced and unvoiced parts. The rate of the coder can be switched from 2.4kb/s, which operates on 20ms frames, to 1.2kb/s by having 60ms frames. Both rates use the same analysis and synthesis building blocks over 20ms. Reliable pitch estimation and very elaborate voiced/unvoiced mixture determination algorithms render the algorithm robust to background noise. However in order to communicate in very severe noisy conditions a noise pre-processor has been integrated within the speech encoder. Owing to robust quantisation methods and careful encoding/index assignment algorithms the coder maintains its good quality with 1% random bit errors and 3% random frame (20ms) erasures.