An improved model for speech excitation using time-frequency characterisation

This paper introduces a new method for approximating the excitation signal in a linear predictive coding (LPC) model over both time and frequency domains. In the frequency domain, a new two-band excitation model is proposed for decomposing the signal spectrum into two periodic and non-periodic components using the smoothed Wigner-Ville (SWV) distribution. This model provides high resolution in the time-frequency plane which gives accurate information about pitch and voicing. The periodic component is then analysed with a new "multi-pulse" like technique to approximate the signal in the time domain through repeating a representative pattern across each frame. It is shown that combining the spectral and the temporal models improves significantly over the performance of our earlier model for the excitation and gives a good approximation to the signal for most speech sounds. Evaluation results show that the proposed coding method operating at 2400 b/s outperforms current standard coders at similar rate.