The transmission of the spectral information requires a major part of the total bit rate in today's medium-to-low bit rate speech coders. The speech spectrum is relatively smooth for a much longer period than the update rate of the spectral information for many speech sounds. A method for utilizing this is by variable frame rate segment quantization which as a first step identifies steady state portions of the speech signal and then represents each steady state segment by a simple approximation. In the present paper we show that segment quantization can be applied to reduce the bit rate necessary for transmitting the speech spectral information by a factor of two without compromising the total spectral distortion. As an example, using a simple scalar quantizer with 40 bits/segment, an average bit rate of 22.6 bits/frame resulted in a average spectral distortion of 1.06 dB. Using a more sophisticated quantizer allow for reducing the bit rate without increasing the spectral distortion.
Bibliographic reference. Svendsen, Torbjörn (1993): "Efficient quantization of speech spectral information", In EUROSPEECH'93, 1143-1146.