Audio Declipping

Clipping, also known as saturation, is a comon phenomenon leading to sometimes seriously distorted audio recordings. Declipping consists in performing the inverse process, to restore saturated audio recordings and improve their quality.

Original signal

Clipped signal

Declipped signal using A-SPADE

SPADE, the SParse Audio DEclipper

SPADE is a declipping algorithm developed by PANAMA (a joint project-team between INRIA and CNRS). It is based on on the expression of declipping as a linear inverse problem and the use of analysis (A-SPADE) or synthesis (S-SPADE) sparse regularization in the time-frequency domain.
To the best of our knowledge SPADE achieves state-of-the-art declipping quality.

Related publications

• S. Kitic, N. Bertin and R. Gribonval. Sparsity and cosparsity for audio declipping: a flexible non-convex approach. In Latent Variable Analysis and Signal Separation, Liberec, 2015.
• S. Kitic, N. Bertin and R. Gribonval. Audio Declipping by Cosparse Hard Thresholding. In iTwist – 2nd international – Traveling Workshop on Interactions between Sparse models and Technology, Namur, 2014.

A good way to get an idea of the declipping performance of SPADE is to listen to some examples.

Results for different input clipping levels

Click on the examples below to compare A-SPADE with another recent algorithm, Consistent IHT, as well as with the Adobe Audition DeClipper (TM) software.

Two example audio tracks are clipped to prescribed SDR (Signal to Distortion Ratio) values and processed by the different algorithms. The original (unclipped) tracks and the clipped versions are given as references along with the achieved declipping results.

Instrumental sample	Pop song sample
Original	Original

INPUT SDR	1 dB	3 dB	5 dB	7 dB	10 dB	INPUT SDR	1 dB	3 dB	5 dB	7 dB	10 dB
Clipped						Clipped
Adobe Audition						Adobe Audition
Consistent IHT						Consistent IHT
A-SPADE						A-SPADE

Results for different types of music samples

A second demonstration compares the declipping performance of three algorithms: Consistent IHT, A-SPADE and S-SPADE).

Ten example audio tracks are clipped to 3dB SDR and processed by the different algorithms. The original (unclipped) tracks and the clipped versions are given as references along with the processing results achieved using another commercial software.

Audio examples	1	2	3	4	5	6	7	8	9	10
Original
Clipped
Commercial software
Consistent IHT
A-SPADE
S-SPADE

To test A-SPADE, you can either download its Matlab implementation that allows reproducing the experiments of the paper A-SPADE, or use Inria’s software demonstration platform, A||go, to test it on the web.

Download the Matlab implementation of A-SPADE

Download Matlab code for tight frame-based A-SPADE:

• For academic research activities only, this Software is freely available under the terms of the following license agreement (download link). To obtain the Software source code, please fill in (items in blue boxes), sign and send two signed copies of this license agreement to:

Inria Rennes Bretagne Atlantique
Attn Ana-Bela LECONTE
Campus de Beaulieu
35042 Rennes Cedex
France
or to stip-rba@inria.fr (for the scanned signed license agreement)

• For all other uses, the software is available under an Inria commercial license. Please contact stip-rba@inria.fr

Warning: this software version is limited to 64 bits per samples natively clipped .wav files (input files have to be clipped in 64 bits PCM and not transcoded from lossy format for this software version). This limitation is removed in the online version presented below.

Test A-SPADE from your web browser

To test the online demo of A-SPADE on A||go, Inria’s software demonstration platform, you have first to Create an account on the A||go platform if you do not have one already. Then choose between using A-SPADE from the a||go platform (to have all parameters and logs available) or using the light interface on this web page.

This demonstration version handles most audio files formats (.wav, .mp3, .mp4, .m4a, .ogg, .flac) and all common bit depth values.

Limitation: This online demo only processes 30 seconds of the two first channels of the provided clipped input file. If you are interested by other use cases, please use this form to contact us.

Use A-SPADE on a||go platform

• Select the A-SPADE application among others ;
• Process your audio file following the guidelines.

Use the A-SPADE light interface

Declipping process duration: The declipping process can last from a few seconds to a few minutes depending on sampling frequency, chosen preset and number of channels.

User Parameters

Your a||go token (available within your a||go account web page)
Your preset (choose one):	Fastest	Average	High Quality
Upload a clipped file	File: No file selected.

Run A-SPADE (do not refresh this page during declipping process)

Declip

For further information about this technology, improvement suggestions, commercial usage or anything else related to this work, please contact us using following form: