Time-varying linear prediction for speech analysis

In this contribution a time-varying linear prediction is proposed for segment-wise speech analysis. In the conventional linear prediction approach, speech segments are analyzed independently from each other assuming stationary conditions within the frames. However, the articulatory speech production process is a continuous non-stationary process. Considering this fact, a time-varying analysis is treated under the constraints of a continuous time evolution of the predictor coefficients. The coefficients can be determined by a least mean square approach. Analysis results show that the proposed method, especially for short segments, leads to smoother time evolutions of the spectral envelopes in comparison to the common segment-wise time-invariant prediction. Furthermore, the proposed method is more robust in contrast to the covariance method regarding variations of the segmentation. The time-varying analysis approach yields a better estimation of the vocal tract resonances which can be utilized for a variety of applications.