SPEECH REDUNDANCIES AND COMPRESSION (VoIP)

3.3
G.711 provides pure quantized logarithmic compression on a sample basis. The basics on speech redundancies are given in [Bellamy (1991), Nikhil (2000), Kondoz (1999)]. Speech signals have several low-level signals and pauses, which allow for quantization of low- l evel signals with a few bits, whereas operations like VAD make use of the pause periods to improve compression by about 40% to 60%. In speech, sample-to-sample correlations are very high, leading to use of adaptive differential PCM (ADPCM) based compression. At a given time, a speech signal is concentrated in a few frequencies, which results in short time periodicity, and LPC makes use of this short time periodicity. The rate of excitation in the vocal tract is pitch. Pitch periods have long-term correlation that results in more compression. Many compression techniques operate in time domain. Some compression techniques use frequency-domain analysis, which is usually referred to as transform-domain compression.