Here is the code related to the article entitled Phase retrieval with Bregman divergences and application to audio signal recovery. Audio examples of reconstructed signals are available online.
To reproduce the results from the paper (generate the estimated signals and the corresponding figures), run the main.py script. Alternatively, a simple demonstration can be run with demo.py.
The following Python librairies are necessary to run the code properly:
The demo.py script allows you to perform phase retrieval (using the exact magnitude spectrogram) and generate the corresponding time signal using any audio file.
To do so, run:
demo.py -f <filepath> -n <number of iterations> -a <algorithm>.
where default values are respectively datasets/tarantella.wav, 1000 and gradient descent.
Implemented algorithms are gradient descent, ADMM, GLA, FGLA, GLADMM. With gradient descent and ADMM, the loss is the left Kullback-Leibler divergence (d=1).
To reproduce the results from the paper, run:
main.py -s <input SNR in dB> -d <dataset> -n <number of iterations> -f <number of files>.
If the argument -s is not specified, the script will perform phase retrieval from exact spectrograms (Section IV-B in the paper).
Conversely, you can specify the input SNR to perform phase retrieval from modified spectrograms (Section IV-C in the paper).
Default values for the other parameters are respectively speech, 2500, and 10.
If you use any of this code for your research, please cite our paper:
@article{Vial2021jstsp,
author={P.-H. Vial and P. Magron and T. Oberlin and C. F{\'e}votte},
title={Phase recovery with Bregman divergences and application to audio signal recovery},
booktitle={IEEE Journal of Selected Topics in Signal Processing (JSTSP)},
year={2021}
}