Objective evaluation of magnitude and phase only spectrum-based reconstruction of the Speech signal

The aim of this paper is to investigate the effects of window shape and its length on the quality of phase-only and magnitude-only reconstructed speech. Speech signal is reconstructed via Least Square Error Estimation (LSEE) and Overlap Add (OLA) methods from its magnitude-only and phase-only spectra. The effects of using Hamming and rectangular windows have been studied, too. To evaluate the quality of reconstructed speech, we employed three objective measures: Log Likelihood Ratio (LLR), Weighted Spectral Slope (WSS) and PESQ. Results show that in case of using Hamming window, phase-only reconstructed speech surpasses its magnitude-only counterpart in frames longer than 256 ms qualitatively. Rectangular window seems to be a better choice in comparison to Hamming window in case of phase-only signal reconstruction, while for magnitude-only reconstructed speech, Hamming window results in better quality and intelligibility. LSEE has better performance than OLA in case of magnitude-only speech reconstruction while OLA outperforms LSEE in case of phase-only speech reconstruction.