The idea
Recommendation engines connect artists who sound alike or share listeners. Basilect connects artists whose creative philosophy is similar but whose genre or scene is not. Two artists are “basilect-connected” if a punk drummer and an ambient producer turn out to describe their process, their constraints, and their reasons for making music in the same way.
It started with a simple observation: LLMs are strangely good at gauging musical taste — a “vibe” — from very little information. I wanted to know if that vibe could be measured instead of just chatted about.
How it works
The signal source is verbatim artist quotes from interviews. No critic voice, no summaries — the artist’s own words only.
- Search — find interview URLs for an artist
- Scrape — fetch article text with trafilatura
- Extract — pull verbatim quotes from the scraped text
- Embed — embed each quote, aggregate per artist via median vector
- Compare — pairwise cosine similarity across all artists
- Discover — rank pairs by similarity
A corpus only counts if it has at least 5 quotes from 3 distinct sources across 2 distinct years, to keep one puff-piece interview from defining an artist. Earlier iterations (preserved in the repo’s tags) used LLM summaries of artist discourse before I concluded that only direct quotes keep the signal honest. Along the way I tried several comparison methods — Jaccard, Spearman, cosine, count-adjusted similarity — and built visualizations of the encoding methods themselves.
Where it led
The hard part turned out to be the data, not the math: interview scraping is full of confounds. A friend forked the project, put real API budget and a frontend behind it, and hit exactly the same walls — which was oddly validating. Chasing the problem properly pushed me into actual recommender-systems literature, and eventually into a Codabench challenge (Music-CRS 2026) run by some of the same researchers whose papers I’d been reading. I got my system to baseline parity before recognizing how deep that field goes.
The honest conclusion: “measure a vibe” is a genuine open research problem. Basilect is on hold, not abandoned — and its sibling project, NTS Artist Map, attacks the same thesis from a completely different signal source.