AI radio, always on. → webplay.io
webplay is a small internet radio station whose music is generated by AI. It streams around the clock: lofi beats for studying, calm music for sleeping, jazz for lounging — with more channels on the way.
If you landed here from the site and are curious how it works, this is the short tour.
Every track you hear was generated by an AI music model running on a home computer — no stock music, no licensed catalogs. But the station isn't a firehose of raw model output. Each channel goes through a curation loop:
- Describe the sound. Each channel has a carefully tuned text description of its target sound — instruments, mood, tempo, texture. Getting this right is most of the work; wording that seems harmless can pull the music in the wrong direction, and some channels were tuned by measuring candidate tracks against a reference recording until they matched.
- Audition short samples. The generator renders a batch of short candidate clips, each from a different random seed (so each has its own musical personality, but stays reproducible).
- A human picks. I listen and choose the seeds with the best feel. The metrics help rank candidates, but ears make the final call.
- Render the real tracks. The chosen seeds are re-rendered as full multi-minute tracks, reviewed again, and approved.
- Build the stream. Approved tracks are crossfaded into a seamless loop, chopped into small segments, and published to edge storage so the player can stream them continuously anywhere in the world.
The website is a single minimal page: a background image per channel, a channel picker, and an audio player that plays each channel's segmented stream. There is no backend server running the radio — the "always on" stream is prerendered audio served from an edge network, so it's fast, cheap, and doesn't fall over.
apps/web— the website: the player page, channel list, and backgrounds.apps/generator— command-line tools that render tracks, run benchmarks, and build/upload the segmented streams.packages/adapters— a common interface over music-generation backends, including each channel's tuned sound profile. Swapping or adding a backend only touches this package.docs— decision records, benchmarks, cost models, and the seed registry that makes every live track reproducible.library— local-only audio workspace (raw renders, approved tracks); not checked into the repo.
The specific models, services, and infrastructure behind each piece are deliberately not detailed here — they change as better options appear. The shape of the pipeline is the stable part: describe, generate, curate by ear, stream.
If webplay keeps you company while you work, study, or fall asleep, you can buy me a coffee. ☕