Objective evaluation of phase and magnitude only reconstructed speech: New considerations

The aim of this paper is to make some improvements in the synthesis part of Analysis-Modification-Synthesis (AMS) framework to reconstruct speech with higher quality. In previous works based on this framework the number of iterations was fixed on 100. This number of iterations was not chosen in accordance with a specific measure. In this paper, we monitored the quality of reconstructed speech per iteration and through a heuristic criterion, controlled the number of iterations. Our results showed that the required number of iterations depends on frame duration, applied window and whether the speech signal is being reconstructed from magnitude or phase spectrum. In case of phase-only speech reconstruction we need larger number of iterations. By applying rectangular window, the phase-only reconstructed speech surpasses its magnitude-only counterpart in frames longer than 64 ms, qualitatively. However, in case of using Hamming window, depending on applying LSEE or OAM, phase-only reconstructed speech will have better quality than its magnitude-only counterpart in frames longer than 256 ms and 128 ms, respectively.