Have you met AI DJ StructFreak? 🤖

@cyrusasfa 💖

@is_s_yun @ISMIRConf awesome bass

Anyone heading to #ISMIR2023?🎶 DM me if you'd like to hang out! Or come say hi during the LBD session – I'll be presenting AI DJ StructFreak🤖 and All-In-One Music Structure Analyzer🎧

With awesome advisor @juhan_nam ✨

TRAINING CODE for All-In-One Music Structure Analyzer is added!

@kjw_audio In fact, the waveforms are 3D and the camera is looking them at the top of them 😂

@kjw_audio Thank you! I guess you are asking about Music Dissector. Svelte for frontend framework. Web audio API for audio playback. For the waveform rendering, I tried HTML canvas but the speed wasn’t satisfying so I used Three.js to enable GPU rendering.

@DasaemJ Thank you very much! 💖

@juan_s_gomez Thank you very much!

@lixiang62574589 I don’t have a plan to implement a real-time version for now. Thanks for the question!

@zhang_zhedong Thanks! Yeah that's why I also implemented this feature :)

@nsthorat Thank you! Actually, I've been studying this for an AI DJ. However, the training dataset (Harmonix) isn't tailored for EDM. So, my colleagues and I are building an EDM dataset using Rekordbox lol. We hope to release it along with the feature you mentioned!

@syao_ming Well, to my knowledge, it's the first time. However, it's a commonly used technique, especially in similar areas like beat tracking. So, it feels a bit shameful to say "the first time" haha.

@syao_ming Thanks! It’s for capturing long-term dependency of segments. But actually beat/downbeat tracking models already use similar size of receptive fields lol such as Beat Transformer

A screenshot of the interactive demo 🎶

@fdelamuerte Thank you very much! 🤩

@gazorp5 Thanks! Elaborated evaluation may needed, but here are some examples of live recordings (not perfect): I think it works if the beats are strong enough.