On variable-scale piecewise stationary spectral analysis of speech signals for ASR