audiolla
audiolla is one Docker image that wraps a pile of open-source audio tools (Demucs, UVR, matchering, pedalboard, librosa, SoX, ffmpeg, basic-pitch, DeepFilterNet, silero-vad, pyannote, CLAP, AST, fluidsynth and five text-to-audio models) behind a single HTTP API on port 8000. You stage a file, POST a JSON body, and get audio, MIDI, images or JSON back; LLM agents get the same tools over MCP at /v1/mcp.
There is no account, no cloud and no per-minute bill. It runs on your machine, on a CPU or an NVIDIA GPU.
Install
# CPU
docker run --rm -it \
-v $HOME/.audiolla-data:/data \
-p 127.0.0.1:8000:8000 \
psyb0t/audiolla:latest
# NVIDIA GPU
docker run --rm -it --gpus all \
-v $HOME/.audiolla-data:/data \
-p 127.0.0.1:8000:8000 \
psyb0t/audiolla:latest-cuda
Then check it is up:
curl http://localhost:8000/healthz
The getting started page walks through the first real request.
Where to read
| You want to | Read |
|---|---|
| Run your first request end to end | Getting started |
| Pick an image, mount data, use a GPU, upgrade | Deployment |
| Look up an environment variable | Configuration |
| Upload, download and expire files, or use S3-style URLs | Files and remote URLs |
| Split stems, remix, remove reverb, echo, noise or hum | Separation and restoration |
| Master, compress, EQ, pan, add reverb or effects | Mastering, dynamics and effects |
| Get BPM, key, chords, loudness, tags, embeddings | Analysis |
| Enhance speech, detect voice, find speakers | Speech |
| Trim, mix, concat, stretch, pitch, convert, tag, visualize | Editing, time and pitch |
| Generate music or sound effects from a prompt | Music and SFX generation |
| Transcribe audio to MIDI, compose, render MIDI | MIDI |
| Chain steps server-side or run many ops at once | Presets, pipelines and batch |
| Run long work in the background with webhooks | Async jobs and webhooks |
| See every engine, what image it ships in, its license | Engines |
| Read every endpoint and field | REST API reference |
| Connect an LLM agent | MCP server and Agent integrations |
| Put it on a network safely | Security |
| Fix an error | Troubleshooting |
| Understand how the pieces fit | Architecture |
| Build, test or change the code | Development |
Images
| Tag | Device | Engines |
|---|---|---|
psyb0t/audiolla:latest, psyb0t/audiolla:vX.Y.Z | CPU | 27 |
psyb0t/audiolla:latest-cuda, psyb0t/audiolla:vX.Y.Z-cuda | NVIDIA GPU (CUDA 12.4 build of Torch) | 35 |
Both are linux/amd64. The CUDA image adds the larger Demucs variants and the five generators; the engines page has the full list.
License
audiolla itself is WTFPL. The images bundle third-party code under its own licenses, including GPL v3 (matchering, pedalboard); THIRD_PARTY.md lists them. Model weights are not in the images: engines download them at startup or first use, under each model’s own license, and some of those are non-commercial (MusicGen). The engines page has the per-model terms.