Demucs pulls one part out of a mix (vocals, drums, bass, guitar or piano) for pennies.
Demucs is the source-separation model that became the standard for taking a mix apart. In Hiroska it works one stem at a time: choose the part you want and that is the file you get back, ready for an audio lane.
At editor lengths it is close to free; a three-minute track costs a fraction of a cent. Lift an acapella for a montage, drop the vocal to use a song as a bed, or isolate drums to cut picture to.
The result is a normal asset, so cut it like any other clip
What it costs
About 2 tokens ($0.02) per 10 s of audio; the exact price depends on the length of what you upload.
Settings
Tokens
Price
10 s of audio
2
$0.02
30 s of audio
5
$0.05
60 s of audio
9
$0.09
Prices come from the live catalog at $0.0100 per token, the pay-as-you-go rate; subscriptions include tokens at a lower rate. The exact price is shown before you generate, and a failed generation is refunded.
Questions
Does Demucs return all the stems at once?
You choose one per run, and that is what lands in your Assets. Running it again for another stem costs the same negligible amount and keeps each result a single clean asset.
Can Demucs remove the vocal from a song?
Yes. Pick “everything else” and you get the backing without the lead vocal, which is the usual way to turn a track into a bed for narration.
How good is Demucs separation?
Demucs is the reference-quality open model for this, and the six-stem variant is what Hiroska runs. For anything you can describe in words rather than pick from a list, the prompt-driven separator is the other option on the same shelf.