An incremental algorithm for signal reconstruction from short-time fourier transform magnitude

We present an algorithm for reconstructing a time-domain signal from the magnitude of a short-time Fourier transform (STFT). In contrast to existing algorithms based on alternating projections, we offer a novel approach involving numerical root-finding combined with explicit smoothness assumptions. Our technique produces high-quality reconstructions that have lower signal-to-noise ratios when compared to other existing algorithms. If there is little redundancy in the given STFT, in particular, the algorithm can produce signals which also sound significantly better perceptually, as compared to existing work.