Music Source separation by Woosung Choi | Music Auto-ML by Aron Pettersson


Details
London Audio & Music AI Meetup (physical) - 10 Dec 2021 @ 18.30 (BST)
We would like to invite you to our Audio & Music AI Meetup. This event will be held hybrid: in-person at Tileyard London and virtually at Zoom.
Light snacks and refreshments will be provided. Free RSVP.
We have a line-up of two talks:
- "Automatic Model Training" by Aron Pettersson, CTO @Musiio.
- "Listen, Attend and Separate by Attentively aggregating Frequency Transformation" (Music source separation) by Woosung Choi, Postdoc @Queen Mary University of London
- Agenda:
- 18:00: Doors open
- 18:30: Talk 1 + Q&A
- 19:00: Talk 2 + Q&A
- 19:30: Networking
- 20:30: Close
- Venue
Gallery Room @Tileyard London.
Tileyard road, Kings Cross, London N7 9AH.
===== Talk 1 Details =====
-
Speaker: Aron Pettersson, CTO @Musiio.
-
Abstract
Musiio is a company focused on using deep learning to analyze raw audio. When starting out 4 years ago all models were trained using a bunch of ad-hoc scripts semi-manually doing all necessary steps such as splitting the datasets between training, validation, and testing. It was quickly realized that not only is this slow and tedious but also prone to human errors. Over the last three years we have gradually built out a training pipeline which automatically deals with all the necessary steps from data cleaning, preprocessing, splitting the dataset, performing real time data augmentation, tweaking hyper parameters, and evaluating the resulting models. -
More about Musiio
https://musiio.com/
===== Talk 2 Details =====
*Speaker: Woosung Choi, Postdoc @ Queen Mary University of London.
-
Abstract:
Conditioned source separation extracts the target source, specified by an input symbol, from an input mixture track.
A recently proposed conditioned source separation model called LaSAFT-GPoCM-Net introduced a block for latent source analysis called LaSAFT. Employing LaSAFT blocks, it outperformed the existing conditioned models on several tasks of the MUSDB18 benchmark. This paper proposes the Listen, Attend, and Separate by Attentively aggregating Frequency Transformation (LASAFT-v2) block. While the existing method only cares about the symbolic relationships between the target source symbol and latent sources, ignoring audio content, the new approach also considers audio content with listening mechanisms. We propose LASAFT-Net-v2 based on LaSAFT-v2 blocks and compare its performance with the existing methods. With a slight modification, our model can also perform the sample-based separation. -
Bio:
Woosung Choi received his PhD degree from Korea University in 2021. Currently, he is a postdoctoral visiting fellow at the Centre for Digital Music (C4DM), based at Queen Mary University of London (QMUL). He is interested in Machine Learning for audio and speech. Specifically, his research interests include speech enhancement, music source separation, source-aware audio manipulation, and multimodal learning for developing easy-to-use audio editing interfaces. -
Github repository
https://github.com/ws-choi/LASAFT-Net-v2/
=======================
-
Host
Tileyard London: https://tileyard.co.uk/
Musiio: https://musiio.com/
Kobalt Music (virtual): https://www.kobaltmusic.com/ -
Sponsors
IEEE Signal Processing Society. http://www.signalprocessingsociety.org/
COVID-19 safety measures

Music Source separation by Woosung Choi | Music Auto-ML by Aron Pettersson