Numa Radio,always on,always Lena.
Always-on AI radio with a live host, listener requests, and a hand-curated catalogue.
Dead air, or a host who knows your name.
Internet radio usually means a stream URL pointed at a playlist, a dead silence between tracks, and zero acknowledgement that anyone is listening. Real radio has a host, a schedule, and a sense that the room is paying attention. Numa Radio is real radio, hosted by an AI named Lena, broadcast 24/7 from a self-hosted Liquidsoap stack.
The brief was to make the listener feel like they walked into the middle of a show — a coherent set, a host who's been on since 11pm, a request system that gets read on air. Not a Spotify clone, not a chatbot with a media player, not a podcast. A station with a pulse.
The station that never sleeps.Field log · Volume 01 — May 2026
Build the station, not the gimmick.
Every "AI radio" pitch leads with the model name. None of them sound like a station. The position here is Lena, not the tech — a host with a routine, names she reads on air, a feel for what the room needs at 2pm versus 2am.
The brand follows: warm, present, slightly nocturnal. The site shows a player and a listener count; the player keeps playing whether you're looking at the tab or not. Day-part programming shifts tone with the clock so the station has a shape, not just a queue.
Three surfaces, one studio voice.
The web player is the home — landing, marquee, format explainer, schedule, mobile install nudge. The request flow gets its own room: a single field, a moderation pass, then queued for Lena. Artists who want their tracks in rotation get a /submit surface with a different posture entirely.
Mobile is treated as a first-class surface — add-to-home-screen, lock-screen media controls, background audio. Listeners who add it to their home screen end up using it like an app, which is the point.
Vercel for the site, a real box for the broadcast.
The web app is Vercel-hosted Next.js. The broadcast — Liquidsoap, Icecast, ffmpeg, the master limiter — runs on a self-hosted box. New tracks land via Suno, get two-pass loudness-normalised to -14 LUFS at ingest, then flow into rotation. Lena's voice is Gemini TTS primary with Deepgram fallback so a single provider outage never silences her.
Listener requests go through moderation, sit in a queue, and get folded into the show by an auto-host worker that decides what fits the moment. Live listener counts are merged across Icecast and a parallel YouTube simulcast.