Usefulness of phase in speech processing
Kuldip K. Paliwal · 2003
Abstract. It is a common belief in the speech community that the short-time phase spectrum plays very little (or, no) role in human perception tasks as well as in automatic speech recognition systems. In this paper, the usefulness of phase information is explored in human speech perception as well as in automatic speech recognition. Through human perception experiments, it is shown that the short-time phase spectrum (with window size of 32 ms) contributes to speech intelligibility as much as the corresponding power spectrum. A representation based on frequencies of the speech signal derived from its short-time phase is developed and is found to be as good as a cepstral representation (derived from power spectrum) for automatic speech recognition.