First International Conference on Spoken Language Processing (ICSLP 90)
The choice an algorithm of speech coding is very important to achieve high quality speech at low bit rates. Speech can be modeled using LPC and Sinusoidal Transform Coding (STC). In LPC, it leads to CELP type coders [l]. In CELP, during vector quantization of the excitation, all components are matched as a single vector. This produces background noise and hence roughness below 4.8 kbits/s. In STC , on the other hand, the model parameters (phase and frequency) are very sensitive to quantization errors. This affects the performance of this system under channel errors even though it produces high quality speech at low bit rates. In our previous work, we used sine wave components to represent the CELP excitation  and LPC residual waveform  which both are capable of synthesizing speech without the artifacts common to model-based speech system. In this paper, we present the sine wave excited linear prediction (SWELP) speech model which has been found to be robust in the presence of quantization noise in speech. These characteristics make the model particularly useful in the development of high quality speech coding system at low bit rates.
Bibliographic reference. Yeldener, Suat / Kondoz, Ahmet M. / Evans, Barry G. (1990): "Sine wave excited linear predictive coding of speech", In ICSLP-1990, 93-96.