A new hybrid CELP-harmonics speech coder at low bit rates

This paper describes a new speech coding system working at a bit rate of 2400 bps. Like other coders at this bit rate, the BTC-S24 is based on a linear prediction (LP) model that uses a bimodal voicing decision in order to make an efficient coding of voiced and unvoiced frames. Such coders have proved to produce a good intelligibility of the synthesized voice but often at a quality that is inadequate for network communication applications like internet telephony. For this raison, some features have been added to the coding system in order to improve the perceptual quality of the produced voice without introducing high complexity. These features are implemented by the way of an hybrid coding procedure based on a pitch harmonics reconstruction for voiced frames and a CELP like algorithm for unvoiced frames. A voiced proportional ratio is also applied in voiced frames to take into account the proportion of voiced and unvoiced energies in several frequency bands. This coder has been tested by different people and results show that the perceptual quality is greatly improved against the original LPC-10.