I travelled to Seoul, Korea for ICASSP 2024 to see Xinlei Niu present our paper SoundLoCD: An Efficient Conditional Discrete Contrastive Latent Diffusion Model for Text-to-Sound Generation.
I hadn’t travelled for a conference for a few years and ICASSP is a big one so I took the opportunity to learn about the cutting edge of AI and signal processing while getting to see Xinlei’s presentation in person.
ICASSP is a really exciting community because, when you think about it, a lot of the interesting data we want to interact with in the world is a signal (of some kind or another). Given the scale of ICASSP I found it almost impossible to navigate the many parallel sessions effectively. I focussed on the plenary sessions which were really fascinating and the (many MANY) posters. I tend to find poster sessions more memorable than talks as I can filter through to the topics I’m focussed on (sound and music) and have quick discussions with the authors.
Filtering through to find many music posters, I learned a lot and I can make a few high level comments:
- genAI for sound and music is popular but not ubiquitous, there is strong work but coming from a few focussed labs, groups and companies.
- traditional industry names in sound/audio research (e.g., Sony, Bose, Dolby) are present along with the big tech interests (Meta, Apple, ByteDance) as collaborators on many of the audio projects.
Interest and skills in sound and audio seem to be a bit sparser than traditional video/image AI research. I suspect that some of the industry players with deep pockets and/or music publishing may also be bringing data into the equation but this is speculation.
Outside of the conference I had a chance to get some street photography, coffee drinking, and music enjoying in. Last time I was in Seoul was 2013 so I was ready to enjoy returning!
