On the Importance of Pre-emphasis and Window Shape in Phase-Based Speech Recognition

This paper aims at investigating the potentials of the phase spectrum in automatic speech recognition (ASR). We show that speech phase spectrum could potentially provide features with high discriminability and robustness. Out of such belief and to realize a higher portion of the phase spectrum potentials, we propose two simple amendments in two common blocks in feature extraction, namely pre-emphasis and windowing, without changing the workflow of the algorithms. Recognition tests over Aurora 2 indicate up to 11.2% and 14.7% performance improvement in average in the presence of both additive and convolutional noises for phase-based MODGDF and CGDF features, respectively. It proves the high potentials of the phase spectrum in robust ASR.