Tutorial · Tutorial · Migration

A Real Demucs Alternative: Same Quality, No GPU

HTDemucs is the open-source quality king — and a small infrastructure project to run in production. If your goal is the output, not the DevOps, this tutorial swaps the GPU setup for an API call.

DemucsVocal separationMigration

Swap Demucs for the API

Step 1 — What you're replacing

A typical self-hosted Demucs setup is: NVIDIA GPU machine, CUDA + PyTorch environment, model checkpoints, a job queue, retry logic, monitoring, and periodic retraining/update work. The output is 4 stems: vocals, drums, bass, other.

Step 2 — Get the API key

Sign up (free tier included)
Copy the key from the dashboard
pip install requests (or the mlslabs SDK)

Step 3 — One song, same quality

Python

import requests, time

API = "https://api.mlslabs.io/v1"
HEADERS = {"X-API-Key": "YOUR_API_KEY"}

job = requests.post(
    API + "/vocal-separator/jobs",
    headers=HEADERS,
    json={
        "input": "s3://bucket/song.mp3",
        "output": "s3://bucket/stems/",
        "mode": "music",      # HTDemucs-family model
        "tracks": ["vocals", "drums", "bass", "other"],
    },
).json()
job_id = job["job_id"]

while True:
    st = requests.get(API + "/vocal-separator/jobs/" + job_id, headers=HEADERS).json()
    if st["status"] == "done":
        print(st["output"])
        break
    if st["status"] == "failed":
        raise SystemExit("failed — credits auto-refunded")
    time.sleep(2)

Step 4 — Bring your batch

The same call scales to a library: submit N jobs, get N webhooks. That is the part self-hosted Demucs leaves to you — queue, retries, monitoring — already built.

Migration checklist

Note: Quality is not 'as good as Demucs' — it is the same model family, tuned per mode. Run one A/B on your own track before migrating.

Replace the CLI invocation with the API call
Point output URLs to your object storage
Wire webhooks into your existing job tracker
Delete the GPU box (or repurpose it)

FAQ

Common questions

Is the API quality really comparable to HTDemucs?

Our music mode is trained in the same model family, with per-mode tuning. A/B test on your own audio — the free tier is there for exactly that.

What about on-prem data rules?

The hosted API moves media to object storage. If data must stay on-prem, self-hosted Demucs remains the right answer.

Can I keep my existing queue and just swap the worker?

Yes — the API is a drop-in worker: submit job, await webhook, fetch output. Most teams replace their custom worker in an afternoon.

Keep exploring

A Real Demucs Alternative: Same Quality, No GPU

Step 1 — What you're replacing

Step 2 — Get the API key

Step 3 — One song, same quality

Step 4 — Bring your batch

Migration checklist

Common questions

Related guides

Vocal Separator API

Spleeter vs Demucs

Vocal Separation API vs Demucs