Show HN: Field* – Explore a map of 1.8M songs based on audio similarity

[Barneyhill]

This project builds upon cosine.club - an amazing music similarity search engine posted the other day (https://news.ycombinator.com/item?id=40072062) which works by comparing audio embeddings between songs with this new model from essentia/upf https://essentia.upf.edu/models.html#discogs-effnet.

I wondered what these embeddings would look like when visualised (using UMAP). Would genre "clusters" pop up? Would similar genres appear next to each other in the embedding space? It turns out they do!

I've ended up using this a lot for discovering new music - there's something really useful about having discovery unbiased by song popularity - hope you enjoy!

Also if you're interested in how ai is impacting music beyond uninspired "song generators" we're working on a series of open-source ai/music experiments at https://vroomai.com/

[mtmail]

Sadly the full desktop dataset (100MB) didn't load for me, I waited 10 minutes on fast broadband connection. The mobile dataset (20MB) stopped loading after 3MB. The file seems fine, I can download it with curl, but inside the browser (Firefox) it seems to timeout.

[Barneyhill]

Ugh that's so annoying... Was originally hosted on a cloudflare static site - now switched to R2 bucket storage. That might help things?