Preferred frame length for the short-time magnitude spectrum on speech intelligibility and speech quality

In many speech processing applications, the short-time magnitude spectrum is considered important and believed to contain most of the intelligibility. In this paper, magnitude-only signals were generated only with short-time magnitude spectra by randomizing the short-time phase spectra of the original speeches. A sentence intelligibility test was implemented to study the speech intelligibility of magnitude-only signals as a function of frame length. Moreover, an objective speech quality measure was used to calculate the quality of magnitude-only signals. The results show that the frame length of 25-64 ms is the optimum choice for both high speech intelligibility and acceptable speech quality, when the short-time magnitude spectrum is only used to reconstruct the speech.