On optimal modelling of speech spectral transitions

In this paper, we propose an optimal spectral transition modelling technique for speech. The proposed technique optimizes the spectral interpolation trajectory by minimizing the mean-square-error of spectral parameters on a frame-by-frame basis. The performance of the proposed techniques is compared with that of two spectral interpolation techniques, namely the linear interpolation and the Gaussian interpolation, reported in literature. Line spectral frequencies are used as the short-term spectral parameter representation of the speech signal. The regions between maximally stable (stationary) frames in the spectral parameter sequence are identified as the regions of spectral transitions. Numerical results show that both linear and Gaussian interpolation techniques have similar modelling performance in terms of average spectral distortion. The proposed optimal technique shows an improved modelling accuracy in terms of average spectral distortion (up to 1 dB improvement), in comparison to that of the linear and Gaussian techniques. The proposed technique can be useful for speech processing applications such as coding and recognition.