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