STFT Phase Reconstruction for Speech Enhancement

The algorithm estimates the phase spectrum of the underlying speech signal. The voiced part of the speech is modeled using the harmonic model. As a prerequisite step, the fundamental frequencies of the voiced speech are estimated. The phase reconstruction algorithm makes it possible to enhance speech using only the fundamental frequencies and the noisy signal.

I later improved upon this algorithm by using pitch-synchronous window size for STFT instead of constant window size. This made the estimation of phase spectrum faster and more accurate. Visit the project’s page to see how it works.

The code is available on GitHub.