Find the exact date and zoom link in our announcements on the Music Reading Group Google group: groups.google.com/g/music_readin. Subscribe to stay updated! Check out previous talks on our YouTube channel (youtube.com/@musicaireadin) in the meantime.
Follow
Click to Follow wuyusongwys
Yusong Wu
@wuyusongwys
1st year PhD student at Mila & University of Montreal.
Yusong Wu’s Tweets
I will be co-presenting a tutorial with at #ISMIR 2022 on **Designing Controllable Synthesis System for Musical Signals**! The tutorial is happening on Dec. 4th, from 9:00AM - 1:00PM Indian time: ismir2022.ismir.net/program/tutori.
2
33
Congratulations Ethan! Ethan has done great works in MIR and Machine Learning during his remarkable PhD time. I always enjoy working with Ethan, and I learned a lot from him!
Quote Tweet
Hey y'all, my dissertation was quietly put up online a few weeks ago and I wanted to highlight some parts of it that I was particularly proud of... 1/10
Show this thread
read image description
ALT
3
Large-scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmentation
abs: arxiv.org/abs/2211.06687
github: github.com/LAION-AI/CLAP/
1
34
132
We would also like to express our huge appreciation to Christoph Schuhmann, Richard Vencu, Irina Rish , Romain Beaumon, Emad Mostaque , as this project would not be possible without them.
6/6
3
8
Show this thread
For more details, please check out our paper and code! This work is done with Ke Chen (), Tianyu Zhang (), Yuchen Hui (), Taylor Berg-Kirkpatrick, and Shlomo Dubnov, collaborating with .
5/6
1
1
10
Show this thread
For the contrastive learning model we build a contrastive learning model with HTSAT audio encoder and Roberta text encoder. With the large-scale dataset and input fusion, our performs well on text-audio retrieval tasks, as well as down-stream tasks.
4/6
1
10
Show this thread
The audio files collected from the internet have variable duration, making it insufficient to learn with a fixed-length encoding model. We adapted feature fusion to the audio encoder to enable the audio encoder to have a fixed compute given any input length.
3/6
1
9
Show this thread
Audio data is often more expensive to collect and annotate, which makes the audio dataset often smaller than the image counterparts. In this work, we announce LAION-Audio-630K, 630k text-audio pairs collected from the internet.
2/6
1
1
17
Show this thread
We are glad to announce our new work collaborating with LAION.ai: a large-scale text-audio contrastive learning model with a large-scale text-audio dataset!
Paper: arxiv.org/abs/2211.06687
Code & Model: github.com/LAION-AI/CLAP/
Dataset: github.com/LAION-AI/audio
1/6
15
154
639
Show this thread
Code for Chamber Ensemble Generator:
2
14
Show this thread
What’s exciting about the CEG is that it uses a set of structured generative models which provide annotations for many music machine learning applications like automatic music transcription, multi-f0 estimation, source separation, performance analysis, and more!
🧵
1
3
Show this thread
We trained an automatic music transcription model MT3 on URMP and CocoChorales. Compared to training on only URMP that our data generator trains on, we get big improvement. This suggests using generate models to generate data helps task performance on low-resource dataset.
🧵
1
4
Show this thread
As its name suggests, the Chamber Ensemble Generator (or CEG) can generate performances of chamber ensembles playing in the style of four-part Bach chorales. We call the generated dataset CocoChorales Dataset, and it has 240,000 examples totaling over 1,400 hours of mixture data.
1
5
Show this thread
Our answer: the model itself!
We combined two structured generative models–a note generation model, Coconet, and a notes-to-audio generative synthesis model, MIDI-DDSP–into a system we call the Chamber Ensemble Generator.
🧵
1
1
11
Show this thread
Modern machine learning systems require large quantities of *annotated* data. With music systems, getting annotations requires tedious work by expert musicians. How can we easily annotate hundreds of thousands of examples to make enough data to train a machine learning system?
🧵
1
12
Show this thread
We’re happy to announce the Chamber Ensemble Generator and CocoChorales Dataset!
We use generative models to create a massive, highly-annotated, and realistic-sounding dataset of chorale performances.
Blog & Samples🔊: g.co/magenta/ceg-an
Paper📝: arxiv.org/abs/2209.14458
🧵
8
44
200
Show this thread
DDSP-VST! Realtime Neural Audio Synthesis as an audio plugin. Excited to finally release this fun way to explore new sounds.
🔊 Plugin: g.co/magenta/ddsp-v
🚂Train Models: g.co/magenta/train-
🎮Discord: discord.gg/eyzhzMJMx5
📝Blog: magenta.tensorflow.org/ddsp-vst-blog
1/
21
139
478
Show this thread
AISC 2022 Finalist ⚡️ 🤖: 3+i - A to I
The main character is an AI model born from chaos doubting its existence but finally embraces the original intention of humans: Developing intelligent machines to improve human lives.
Listen & vote here: aisongcontest.com/participants-2
1
4
Last, don't forget to vote for our entry if you like our song!
🗳️: voting.amsterdammusiclab.nl/vote/2022/24
1
Show this thread
Check out all entries: aisongcontest.com/participants-2
And to check out and vote for the top-15 finalists:
1
2
Show this thread
Please check out our entry in AI Song Contest 2022! It's about the growth of the AI model from its own perspective🤖
🎤: aisongcontest.com/participants-2
Please vote for us if you like our song!
🗳️: voting.amsterdammusiclab.nl/vote/2022/24
#aisongcontest
1
2
14
Show this thread
It's ICLR time! Come check out 's great work with and others on MIDI-DDSP!
We're honored to be in "Oral 1: AI Applications" today (5pm PT, iclr.cc/virtual/2022/s)
Yusong made a great video that clearly explains the model, so check it out!
1
4
15
Show this thread
The #aisongcontest is back! Register by 15 March. Submit by 1 June.
aisongcontest.com/participate-20
This year, we’ll have mentors to help teams get started!
🌱 To GET a mentor: forms.gle/brTusf72w5T3Ai
💜 To BE a mentor: forms.gle/ubdycK2XCohj3u
Details: aisongcontest.com/mentoring
Quote Tweet
Be part of the AI Song Contest 2022! 
You can register your team until 15 March 
. If you are looking for team members - send us a message!
(In the picture you can see all the wonderful teams and cover art that brought the contest to life last year
) #aisongcontest
You can register your team until 15 March . If you are looking for team members - send us a message!
(In the picture you can see all the wonderful teams and cover art that brought the contest to life last year) #aisongcontest
16
33
Glad to announce our rencent ICLR paper on musical performance generation with detailed control!
Quote Tweet
MIDI-DDSP: Detailed Control of Musical Performance via Hierarchical Modeling. Coming to an ICLR near you as an oral presentation!
Blog: g.co/magenta/midi-d
Examples: midi-ddsp.github.io
Colab: g.co/magenta/midi-d
Code: github.com/magenta/midi-d
1/
Blog: g.co/magenta/midi-d
Examples: midi-ddsp.github.io
Colab: g.co/magenta/midi-d
Code: github.com/magenta/midi-d
1/Show this thread
1
2
31
MIDI-DDSP: Detailed Control of Musical Performance via Hierarchical Modeling 🎶
abs: arxiv.org/abs/2112.09312
github: github.com/magenta/midi-d
1
27
140
Show this thread
Composer turned ML researcher Cheng-Zhi Anna Huang (, ) takes a human-centered approach to ML & creativity, expanding artists' creative reach. Huang also helped create Google’s 1st AI Doodle, the “Bach Doodle”.
3
19
57
Show this thread