The archive · Product Ideas · Product decision · 2023
Maroofy searches 120M+ iTunes songs by sound, not metadata, to find similar music
One developer embedded the audio of ~120M iTunes tracks into a vector space, then let anyone search by song — 753 HN points and a 427-comment listening test.
subtech (author of Maroofy)
What it had to solve
A developer posting as subtech wanted to search any song and hear others that genuinely sound alike — something playlist algorithms built on listening history and popularity rarely deliver, especially outside mainstream genres.
How it works
On 3 February 2023, a developer posting as subtech shipped Maroofy with the simplest possible ask on its homepage: search for any song and it will use the song's audio to find other similar-sounding music. Underneath was an unusual bet — instead of recommending from listening history or genre tags, he trained a custom AI audio model that reads raw audio and outputs embedding vectors, indexed roughly 120M songs from the iTunes catalog, and stored the vectors in a database for semantic search.
The trick that made catalog-scale audio feasible surfaced in the comments: Apple's public iTunes API returns a free preview clip for every track, so the model could be built on about 30 seconds of each song. Commenters noticed the consequence immediately — recommendations match the preview excerpt, so a multi-part track like Bohemian Rhapsody returns neighbours for its middle section, not the whole song.
The launch drew 753 points and 427 comments, and the thread became a public listening test: dozens of users pasted maroofy.com/songs links to stress the model with everything from fusion to Nine Inch Nails, some calling the results 'loads better than Spotify' or 'a feature, not a bug', others hunting for the system's blind spots. The author stayed in the thread, agreed the model was outdated, and promised newer architecture plus upvote/downvote signals.
Why it lands
- Embedding raw audio — not metadata or play counts — made similarity a geometric question, so recommendations could ignore popularity, payola and genre labels entirely.
- Apple's free 30-second preview API gave one developer a legal, programmatic path to most of a 120M-track catalog, turning an impossible data problem into a batch job.
- A single search box invited immediate use, so the launch thread filled with edge-case songs that exposed exactly where audio embeddings fail.
- The author's candour about model limits converted criticism into a public roadmap, keeping a demanding audience engaged instead of defensive.
What it did
The Show HN drew 753 points and 427 comments on 3 Feb 2023; users stress-tested it with dozens of maroofy.com/songs links, called results 'loads better than Spotify' and sometimes gloriously weird (Nine Inch Nails recommending the Muppets); the author answered feedback, calling his model architecture outdated and promising voting and improvements.
What you can take
Scale can be the pitch: embedding a whole catalog made Maroofy's 'sound, not fame' promise concrete, and a plain search box turned commenters into an unpaid evaluation team.
Since then
Maroofy stayed online after launch and grew beyond the demo: later tool listings describe saved favourites, search history, CSV playlist exports and a paid tier, and DJ-discovery articles still recommended it in 2026 as an audio-similarity engine. In the launch thread itself, the author's planned improvements — newer model architecture and explicit user feedback on results — marked the path from the February 2023 prototype toward a maintained service.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe the brief you are staring at, and see who has been given the same problem.
Free account · 3 free questions · no card