An Incremental Algorithm for Signal Reconstruction from
Short-Time Fourier Transform Magnitude
J. Bouvrie and T. Ezzat, to appear at ICSLP 2006.

This webpage provides demonstration audio samples, accompanying the paper as discussed in the "Experiments" section. GL Positive and GL Zero-Mean refer to the Griffin-Lim algorithm when applied to spectrograms derived from positive signals and zero-mean signals respectively. Hop Size specifies the number of samples between the start of adjacent STFT analysis windows. A smaller hop size implies greater redundancy in the STFT, so that the corresponding spectrogram provides more information about the original signal. Each STFT was computed with rectangular (boxcar) windows. Note that the success of this algorithm strongly depends on the use of a rectangular window. If your processing requires the use Hamming/Hanning or other non-rectangular windows to compute the STFT, the methods described herein will probably not give good results. Download the paper here.

Male Speech Sample, 100 sample STFT window size, 200 FFT bins per window Male speech original Algorithm GL Positive GL Zero-Mean Our Algorithm Hop Size 10 10 10 20 20 20 30 30 30 40 40 40 50 50 50 60 60 60 70 70 70 80 80 80 90 90 90

Female Speech Sample, 100 sample STFT window size, 200 FFT bins per window Female speech original Algorithm GL Positive GL Zero-Mean Our Algorithm Hop Size 10 10 10 20 20 20 30 30 30 40 40 40 50 50 50 60 60 60 70 70 70 80 80 80 90 90 90

Drum Loop Sample, 100 sample STFT window size, 200 FFT bins per window Drums original Algorithm GL Positive GL Zero-Mean Our Algorithm Hop Size 10 10 10 20 20 20 30 30 30 40 40 40 50 50 50 60 60 60 70 70 70 80 80 80 90 90 90