Musical Audio Source Separation using Deep Learning
- Submitting institution
-
The University of Huddersfield
: A - Music
- Unit of assessment
- 33 - Music, Drama, Dance, Performing Arts, Film and Screen Studies : A - Music
- Output identifier
- 45
- Type
- T - Other
- DOI
-
-
- Location
- -
- Brief description of type
- Multi-component: Conference Contributions including Contextual Information
- Open access status
- -
- Month
- -
- Year
- 2016
- URL
-
-
- Supplementary information
-
-
- Request cross-referral to
- -
- Output has been delayed by COVID-19
- No
- COVID-19 affected output statement
- -
- Forensic science
- No
- Criminology
- No
- Interdisciplinary
- Yes
- Number of additional authors
-
0
- Research group(s)
-
-
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- This output explores the automatic separation of instrument sounds in musical audio, which empowers new creative applications. Traditionally framed as ‘blind source separation’, in previous approaches programs were adjusted individually for each music track to be extracted. The field transitioned to a more data-driven, supervised approach enabled by the use of deep neural networks (DNNs). The associated increase in quality of the separation paved the way for implementation into commercial applications. This output presents contributions both from a theoretical perspective, concerning separation algorithms, and a more applied point of view, dealing with how such algorithms can be applied to creative workflows. Item 2 was one of the first high-profile journal publications describing the use of DNNs for source separation. Item 1 studied the use of convolutional layers in models to retain contextual information. This research leveraged and contributed to the momentum created in the audio signal processing community around the availability of public datasets, rigorously following the methodology proposed for the SiSEC evaluation campaign. It was presented at the LVA-ICA conference in 2018, along with source code for reproducing the results. Items 3 and 4 explore the use of source separation in music remixing and upmixing. For these creative contributions, DNN-based algorithms were implemented in prototypes that demonstrate their feasibility in real-world applications, including the first ever web audio-based implementation of supervised audio source separation. This research was presented in applied research venues with a mixture of academic and industry participants (Workshop on Intelligent Music Production 2016 and Web Audio Conference in 2017). The code for the web-based remixing was also made publicly available. The relevance to real-world music production workflows has been further established by commercial implementations of DNN-based supervised musical audio separation.
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -