Phase Retrieval: Application to Audio Signal Reconstruction

Abstract

Theoretically, phase retrieval is an efficient method for signal reconstruction given only the magnitude spectrum of the short-term Fourier transform (STFT). This topic has recently regained popularity due to its utility in a variety of applications such as compressive sensing, speech synthesis, speech enhancement, source separation, and so on. As a result, based on an explicit relationship between STFT magnitude and phase, this paper presents an efficient algorithm for audio signal reconstruction using phase retrieval from the STFT magnitude spectrum. The standard metrics in signal reconstruction, such as time domain segmental signal-to-noise ratio (segSNR), time-frequency domain signal-to-error ratio (SER), and cepstrum-related distance measures, such as log-likelihood ratio (LLR), Itakura-Saito distorsion (IS), and cepstrum distance, are used to perform an objective evaluation. The results support the proposed approach's improvement.