Musical Audio Repurposing using Source Separation

Lead Research Organisation: University of Surrey
Department Name: Vision Speech and Signal Proc CVSSP


Delivery of audio has become increasingly complex: originally in single channel (mono) or 2-channel stereo format, now surround sound in "5.1" format (5 main speakers plus one low frequency effects channel) is available in many home cinema systems, and many other multichannel audio formats are available (e.g. 6.1, 7.1, 10.2 and 22.2). In addition, new interactive apps allow users to remix musical audio, changing instrument volumes, and music games allow players to control individual instruments. Content creators therefore have to develop new ways to create and distribute their audio content to allow their content to be played back on these multichannel systems, or remixed by users to suit their own tastes.
However, much audio content is still in legacy formats, mainly 2-channel stereo. We therefore need ways to "repurpose" this legacy audio content, converting these into surround sound or to the separate "stems" needed for remixable audio.
The aim of this project is to develop a new approach to high quality audio repurposing, based on high quality musical audio source separation. To achieve this we will combine new high resolution separation techniques with information such as musical scores, instrument recognition, onset detection, and pitch tracking. Instead of aiming at generic source separation, we will develop algorithms designed to match the separation performance to the final target (upmixing or remixing). In parallel, we will investigate perceptual evaluation measures for source separation, remixing and upmixing, and develop new diagnostic evaluation techniques tailored to measure different aspects of the repurposed outcome.
The outcomes of this project will allow music consumers to enjoy their favourite songs in interactive remixing apps and games, even where the original separate "stems" are not available. It will also allow music companies, broadcasters and sound archive holders to provide high quality upmixed versions of their large archive content, for an increasing generation of listeners with surround sound systems in the home.

Planned Impact

(Non-academic beneficiaries are outlined here and in "Pathways to Impact". For more on academic impact, see "Academic Beneficiaries" and the "Academic Impact" section in the Case for Support.)
Audio researchers in industry will benefit from new methods for upmixing and remixing emerging from the project.
Manufacturers of audio upmixing equipment and plugins, and broadcasters wishing to upmix legacy 2-channel stereo content, will benefit from our new high-quality upmixing methods. Manufacturers of other musical audio effects boxes will benefit from new methods for remixing allowing repurposing of legacy audio content.
Other holders of legacy audio and audiovisual archives, such as the British Library, BFI and regional sound archives, will benefit from the ability to upmix their content for modern audiences becoming increasingly used to surround sound audio.
There is a strong interest amongst both professional and high-end consumer audio users in new methods for unmixing 2-channel stereo content to 5.1 surround sound, leading to a range of upmix (or "unwrap") plugins for systems such as ProTools. These users will benefit from new upmix approaches emerging from this project, either through direct use of research prototypes, or through enhanced software or tools from audio equipment or plugin manufacturers.
Sound artists and composers will benefit from our remixing methods, allowing them to use sounds from mixed audio signals as part of compositions.
Remixing apps are becoming available for mobile devices, allowing users to remix and share audio tracks. Currently these are limited to use tracks where the separated sources are available from the original music label. These remix users and companies would benefit from the ability to remix from the stereo content that they already own.
The staff employed on the project, including postdoctoral research assistants undertaking the research, will gain skills applicable to industrial problems such as advanced digital signal processing, research software development, and evaluation methodologies.


10 25 50
publication icon
Benetos E (2017) Polyphonic Sound Event Tracking Using Linear Dynamical Systems in IEEE/ACM Transactions on Audio, Speech, and Language Processing

publication icon
Cano E (2019) Musical Source Separation: An Introduction in IEEE Signal Processing Magazine

publication icon
Grais E (2017) Two-Stage Single-Channel Audio Source Separation Using Deep Neural Networks in IEEE/ACM Transactions on Audio, Speech, and Language Processing

publication icon
Grais E.M. (2016) Single channel audio source separation using deep neural network ensembles in 140th Audio Engineering Society International Convention 2016, AES 2016

publication icon
Grais E.M. (2018) Combining fully convolutional and recurrent neural networks for single channel audio source separation in 144th Audio Engineering Society Convention 2018

Related Projects

Project Reference Relationship Related To Start End Award Value
EP/L027119/1 01/11/2014 31/12/2014 £887,607
EP/L027119/2 Transfer EP/L027119/1 01/04/2015 31/10/2018 £856,793
Description New methods for audio source separation using deep neural networks.
New methods for remixing and upmixing musical audio using audio source separation.
New insights into perceptual evaluation of audio source separation methods.
Exploitation Route Potential applications for audio upmixing for spatial audio, or remixing in audio production or games or for remixing apps.
Sectors Creative Economy,Culture, Heritage, Museums and Collections

Description Audio-Visual Media Research Platform
Amount £1,577,223 (GBP)
Funding ID EP/P022529/1 
Organisation Engineering and Physical Sciences Research Council (EPSRC) 
Sector Academic/University
Country United Kingdom
Start 08/2017 
End 07/2022
Description H2020-ICT-2015 Audio Commons
Amount € 2,980,000 (EUR)
Funding ID 688382 
Organisation European Commission 
Sector Public
Country European Union (EU)
Start 02/2016 
End 01/2019
Description UK Acoustics Network
Amount £561,807 (GBP)
Funding ID EP/R005001/1 
Organisation Engineering and Physical Sciences Research Council (EPSRC) 
Sector Academic/University
Country United Kingdom
Start 11/2017 
End 11/2020
Title Two-way mixer 
Description TwoWayMixer is a demonstration prototype that showcases the utilization of Deep Neural Networks to audio remixing through source separation, implemented using web audio technology. 
Type Of Technology Software 
Year Produced 2016 
Open Source License? Yes  
Impact The demo allows researchers and practitioners to understand the potential of source separation algorithms for audio remixing, and contributes to the emerging field of Web Audio. It was presented in the Second Web Audio Conference (WAC 2016). 
Title untwist 
Description Untwist is python library for audio source separation. It provides a self-contained object-oriented framework including common source separation algorithms as well as input/output functions, data management utilities and time-frequency transforms. 
Type Of Technology Software 
Year Produced 2016 
Open Source License? Yes  
Impact The development of library has enabled research on audio source separation within the Musical Audio Repurposing using Source Separation project. By releasing as open source, we help researchers in the field of signal processing and source separation to transition to the Python programming language, and encourage good practices in research software engineering. The library was presented at the 9th European Conference on Python in Science (EuroScipy 2016). 
Description AES Convention Milan Workshop and Panel 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Industry/Business
Results and Impact The project organised/chaired a workshop on Audio Repurposing using Source Separation at the AES Convention in Milan, May 2018. Representatives from the project, as well as invited panellists from BBC R&D, Fraunhofer IDMT and Fraunhofer IIS, presented about the state of the art in source separation. This was then followed by a panel discussion about future directions.

The aims of the workshop, building on the idea of using state-of-the-art source separation in the real world, were fourfold:
1. Give an overview of the state of the art in source separation and how it sounds;
2. Discuss and demonstrate current applications of source separation in audio production;
3. Discuss evaluation of source separation in applied contexts;
4. Give perspectives on future research directions and applications of source separation.

The workshop had lots of audio examples and the audience were generally optimistic about the role of source separation in professional audio contexts. The talks sparked interesting questions. We were invited to repeat the workshop at the New York convention in October 2018.
Year(s) Of Engagement Activity 2018