Text prompts
Write what you want in ordinary language — the second speaker, the rain, the bassline. There is no class list to memorise and no fixed set of stems to choose between.
Promptable audio source separation
Describe a sound. Get it on its own track.
SAM Audio Lab pulls a single sound out of a crowded recording from a plain-text prompt — no stem presets, no manual spectral editing. You get the sound you asked for and the residual mix beside it.
Works on audio files and on the soundtrack of a video.
Promptthe acoustic guitar, not the vocal
Target
Residual
Capabilities
One prompt, one clean stem — across speech, music and everyday sound.
Point-and-pick prompts
When a sound is easier to show than to describe, mark it on the video frame where you hear it and the model takes the object you pointed at.
Time-span prompts
Select a stretch of the timeline that contains the sound. A few clean seconds are often all the model needs to identify the target.
Combined prompting
Text, a visual pick and a time range stack together. Two weak cues usually resolve a target that neither one settles on its own.
A single model for every layer
Separation that does not care what kind of sound it is
Try it on your own fileOne model handles the whole mixture instead of a chain of specialised tools. The same prompt interface applies whether the target is a voice, an instrument or a passing car.
Everything
How it works
Three steps, no session setup
Nothing to install and nothing to configure before the first run.
- 01
Bring the recording
Upload an audio file, or a video whose soundtrack you want to work on. The picture is treated as a cue, not as decoration.
- 02
Say what to isolate
Type the target, click it in the frame, or drag over the seconds where it plays. Refine the prompt and run it again as often as you like.
- 03
Take both sides
Download the isolated target and the residual mix as separate tracks, ready to drop straight back into an editor.
Where it gets used
Recordings you cannot go back and capture again
Most audio worth saving was recorded once, in a room nobody controlled.
Podcast and interview cleanup
Take a guest's voice away from the café behind them without flattening the room into a tunnel.
Film and video post
Recover usable dialogue from a location take, or lift an effect out of a mixed track when the original stems are gone.
Music practice and study
Solo one instrument from a recording to learn the part, or mute it and play along with the rest of the take.
Field and nature recording
Pull one bird, one engine or one voice out of a landscape that was never actually quiet.
Captions and accessibility
Feed a cleaner speech track into transcription so captions stop guessing at overlapped words.
Dataset preparation
Build single-source clips out of messy in-the-wild recordings instead of discarding them.
Inside a run
The target and everything else, side by side
Separation is not subtraction. Each run returns the sound you prompted for together with the residual mixture it came from, so nothing is thrown away and either side can be the one you keep.
Promptthe saxophone, not the crowd
Target stem
The prompted sound, on its own track.
Residual mix
Everything the prompt did not select, still in sync.
Original
Untouched, for A/B against both outputs.
An illustrative view of a run. The working interface lives in the tool itself.
FAQ
Questions people ask first
What promptable separation does, and where it gets difficult.
Instead of choosing from a fixed menu of stems — vocals, drums, bass, other — you describe the sound you want in words, or point at it, and that is what gets separated. The set of possible targets is open rather than baked into the model's outputs.
Put the sound you need on its own track
Bring a recording, describe what to isolate, and take both sides away.
Separate a recording
No plugin chain, no stem presets.