Comparison · Open source vs API
Vocal Separation API vs Demucs (HTDemucs): An Honest Comparison
Demucs deserves respect: it is state of the art in open-source music separation, and our music mode is trained in the same model family. The real question is what running it costs you.
What Demucs actually is
HTDemucs (hybrid transformer demucs) is a waveform-spectrogram U-Net with a transformer on top. It is genuinely strong — community benchmarks put it at the top of open-source separation. Running it well is the hard part.
Side by side
| Self-hosted Demucs | mlslabs Vocal Separator API | |
|---|---|---|
| Quality | State of the art | Same model family, per-mode tuning |
| Speed | CPU: 10–15 min per track | Seconds to a couple minutes |
| GPU/VRAM | CUDA + 6–8 GB VRAM minimum | None on your side |
| Ops | Docker, checkpoint mgmt, queue | Zero |
| Batch | You write the queue | Native batch + webhooks |
| Cost | GPU machine 24/7 | Per minute of audio |
When to pick Demucs
If your audio must never leave your environment, or you already run a GPU fleet for other work — self-host Demucs and own the whole stack. It is the right call at scale with existing infra.
When to pick the API
When you want separation as a feature, not a project: an app feature, a pipeline step, a batch job, variable volume. The API turns a GPU project into a function call.
Benchmarks
Keep exploring