APPLADE

An audio declipping method that takes advantages of both sparse optimization and deep learning.

APPLADE (Adjustable Plug-and-PLay Audio DEclipper)

Tomoro Tanaka (Department of Intermedia Art and Science, Waseda University, Tokyo, Japan)

This README file describes the MATLAB codes provided to test, analyze, and evaluate the methods named APPLADE.
APPLADE is an audio declipping method introduced in the following paper

[1] Tomoro Tanaka, Kohei Yatabe, Masahiro Yasuda, and Yasuhiro Oikawa, "APPLADE: Adjustable plug-and-play audio declipper combining DNN with sparse optimization," in IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2022 (accepted).

Requirements

The codes were developed in MATLAB version R2021a and have been tested in R2021a and R2021b.
Some functions rely on

1. MathWorks Toolbox: You are kindly requested to download some of them, such as 'Deep Learning Toolbox' and 'Parallel Computing Toolbox'.

2. Toolboxes available online: These are available online under the MIT license.

• DGTtool
  A simple and user-friendly MATLAB tool for computing the short-time Fourier transform (STFT) and the discrete Gabor transform (DGT). I already installed it so you can easily execute the codes. Plaese refer to https://github.com/KoheiYatabe/DGTtool or its helps for more detailed information.

• calcCanonicalDualWindow.m
  This is a function for generating the canonical dual window. It is from the MATLAB codes that is available in https://doi.org/10/c3qb.
  Please refer to the paper below for more detailed information and other helpful codes.

  [2] Kohei Yatabe, Yoshiki Masuyama, Tsubasa Kusano and Yasuhiro Oikawa, "Representation of complex spectrogram via phase conversion," Acoustical Science and Technology, vol.40, no.3, May 2019. (Open Access)

Data

There are 4 audio data in the folder Dataset/Examples.
They are from LibriSpeech ASR corpus, which is a corpus of English speech sampled at 16kHz. This is freely available under CC BY 4.0 license.
Please refer to the URL above and the paper

[3] V. Panayotov, G. Chen, D. Povey and S. Khudanpur, "Librispeech: An ASR corpus based on public domain audio books," 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206-5210.

for more information about this corpus.

Usage

Execute main.mlx to perform APPLADE. The trained DNN parameters that were used in our experiments are to be used.

• Declipping

  • DGTtool-main contains DGTtool explained above including the license file.
  • Tools contains some functions used in main.mlx and so on.
  • main_APPLADE.mlx is the mainloop of APPLADE.

• Training

  • Train_main.mlx is for training a DNN in your own manner.
  • Models contains model functions to be used as a DNN.
  • Tools contains some functions used in Train_main.mlx and so on. calcCanonicalDualWindow.m is in this folder.
  • modelParameters contains the trained DNN parameters, and your own DNN parameters are also to be in this folder.

License

See the file named LICENSE.pdf.