# ∞ TAPEDECK — Infinite Personal AI Radio, Fully Local

Turn a folder of songs into a radio station that never ends. TAPEDECK is a self-hosted, open-source AI music generation radio built on ComfyUI and MiniMax Music 3 (open weights). It listens to your music library, learns your taste, and generates an endless stream of new songs like yours — but new — on your own GPU. No cloud, no subscription, no telemetry. Just you, one graphics card, and a cassette deck from a cyberpunk 1994.

Drop a folder of hand-picked songs → get a Pandora-style AI radio station for that exact mood, running infinitely, locally.

What it does

• Local AI music generation radio — full songs (vocals or instrumental) generated with MiniMax Music 3 open weights on a single consumer GPU (16 GB VRAM, tested on an RTX 5080).
• Learns your taste from your own library — a listener stack (librosa features + CLAP audio embeddings + optional NVIDIA Music Flamingo captions) distills your collection into taste veins and essence cards.
• Stations = folders. Your whole library is a station. Twelve songs you love at 2 a.m. is a station. Under 24 tracks, no clustering — the folder is the mood.
• A critic with your ears — every generated take is CLAP-scored against your corpus before you hear it; off-taste takes are rejected automatically, with thresholds self-calibrated per vein from your own music.
• The spool — generation is slower than realtime, so a background daemon keeps a reserve of approved takes wound ahead whenever your GPU is idle, and yields instantly when you need the card (games, other AI workloads, a sibling ComfyUI instance).
• Feedback that steers — keep / skip / dislike buttons re-weight what gets generated next. The radio converges on you.
• 90s cassette deck UI — spinning reels, VU meters driven by the actual audio signal, VFD phosphor readouts, tape counter. Works from any browser on your LAN, phone included.
• Drag-and-drop library management — drop songs or folders onto the deck, rescan, exclude tracks from a station's essence non-destructively.

How it works

your music folder
      │  capture (one command / one click)
      ▼
features (BPM, key, energy arc) ──┐
CLAP embeddings ──────────────────┼──► taste veins ──► essence cards
optional AI captions (Flamingo) ──┘         │
                                            ▼
                    caption sampler (+ optional local LLM for lyrics)
                                            │
                                            ▼
              MiniMax Music 3 generation (ComfyUI, your GPU)
                                            │
                                            ▼
                    CLAP critic — sounds like your taste? no → reject
                                            │ yes
                                            ▼
                              the spool → your radio deck

Generation and analysis share one GPU through a duty cycle: captioning yields to the radio when the spool runs low and resumes once it rewinds.

Requirements

• Linux (Windows via WSL2 — see docs/INSTALL.md; native Windows and macOS are not supported), Python 3.12, an NVIDIA GPU with 16 GB VRAM (RTX 4080/5080 class)
• ComfyUI 0.33+ with the MiniMax Music 3 model files (~14 GB, see Comfy-Org/MiniMax-Music-3)
• Optional: any OpenAI-compatible local LLM endpoint (llama.cpp, llama-swap, Ollama) for caption variety and lyric writing — without one, the radio still runs on essence-card seeds
• ~17 GB disk for Music Flamingo, which listens to your library and writes what it hears. This runs by default (--no-captions to skip) because without it a station is described to the generator only by numbers (tempo, brightness, energy) — no genre, no instruments — and takes drift off-style. About 15 s per track, resumable, yields the GPU to the radio.

See docs/INSTALL.md for the full walkthrough.

Quickstart

TAPEDECK has no virtualenv of its own — it is a ComfyUI custom node, and everything runs in the interpreter that runs ComfyUI. Below, that interpreter is $COMFY_PY. Full walkthrough in docs/INSTALL.md.

# 0. ComfyUI 0.33+ installed, Music 3 weights in place (docs/INSTALL.md 1-2)
export COMFY_PY=/path/to/ComfyUI/venv/bin/python
export COMFY_DIR=/path/to/ComfyUI

# 1. clone anywhere, link the deck into ComfyUI
git clone https://github.com/CharlesMod/infinite-tapedeck tapedeck && cd tapedeck
ln -s "$PWD/comfyui_node/music_studio" "$COMFY_DIR/custom_nodes/"
echo "$PWD" > comfyui_node/music_studio/tapedeck_base.txt

# 2. dependencies into that SAME interpreter
$COMFY_PY -m pip install -r requirements.txt

# 3. check the install before it does an hour of work
$COMFY_PY analysis/preflight.py

# 4. put music in ./library, then capture it
#    the AI listening pass runs by default (--no-captions to skip)
$COMFY_PY analysis/import_pipeline.py --station full-library

# 5. run the spool daemon, open the deck, press PLAY
$COMFY_PY radio/tank_daemon.py &
# http://<host>:8188/extensions/music_studio/index.html

FAQ

Is this another AI music generator UI? No — generators make a song when you ask. TAPEDECK is a radio: it models your taste from music you already love, generates continuously in the background, rejects its own misses, and plays an infinite station you never have to prompt.

Does it copy my songs? No. Your library is analyzed (tempo, key, energy, embeddings, optional text descriptions) to steer generation toward a style region — captions never name artists, and the critic compares statistical audio embeddings, not waveforms.

How close does it actually get? It converges on a style region, not a sound-alike. MiniMax Music 3 is conditioned on text: TAPEDECK's listener stack turns your library into words to steer generation, then uses audio embeddings to reject takes that miss. Nothing in the open-weights world today accepts your audio as a conditioning signal, so "in the neighborhood, endlessly" is the honest ceiling. Run the AI listening pass — the difference between a station described in prose and one described in BPM numbers is large.

It rejects everything / the spool never fills. Usually a very tight library (one artist, one album): its own calibrated bar sits higher than generated audio can reach. The critic notices after four takes and switches to a bar learned from what that vein actually produces — the daemon log says corpus 0.9xx out of reach. Raise critic_keep_frac in radio/config.json to make it less picky still.

Why local? Your library is personal. Your taste model doubly so. And a 16 GB GPU that games at night can be a radio station by day.

Keywords: AI music generation, local AI radio, MiniMax Music 3, ComfyUI custom node, open weights music model, personal radio, Pandora alternative, infinite playlist, CLAP embeddings, Music Flamingo, music captioning, self-hosted, offline music AI, text-to-music, taste model, RTX 5080.

Credits

Built on the shoulders of: MiniMax Music 3 (music generation, open weights) · ComfyUI (inference server) · LAION CLAP (audio-text embeddings) · NVIDIA Music Flamingo (music understanding) · librosa.

Respect the licenses of the models you download; Music 3 ships under the MiniMax Community License, Music Flamingo is NVIDIA-noncommercial.

Status

Early release — extracted from a working single-machine install. The pieces run in production daily (the author's living room); the packaged install path is young. Issues and PRs welcome.