Workspace / Record, then separate
Waiting for model
AUDIO–VISUAL SEPARATION

See the voice. Hear what matters.

Separate the speaker you can see from the voices around them.

One voice.
In focus.ONLINE DEMO / 001
01

Visual focus

VISUAL FOCUS
Camera off
Devices off

Give one voice your focus

Enable your camera and mic. Mouth tracking follows automatically.

Mouth detection runs in your browser · Audio and mouth crops are uploaded for separation after recording
Ready for a clearer voiceUp to 60 s · The whole recording is separated on the GPU afterwards
02

Hear the difference

AUDIO MONITOR
16 kHz · MONO · WAV
A

Mixed input

INPUT
−∞ dB
Microphone off00:00
B

Target speaker

ISOLATED
FALCON OUTPUT
The separated voice appears after recording00:00
Your separated audio will appear hereListen or download after separation · Kept only in your browser
A LITTLE GUIDANCE

Bring your target voice into focus.

  1. Enable devices

    Allow camera and microphone access. Keep your mouth clearly visible in the frame.

  2. Follow the mouth automatically

    Auto tracking detects the face and mouth in your browser. The green outline follows the lips. If the face is briefly lost, the last crop is kept until the face returns. Select Reacquire when the target changes, or choose Manual crop; dragging the box or using arrow keys switches to manual mode.

  3. Record, then separate

    Select Start recording, speak or play some sound, then select Finish and separate. Up to 60 seconds; the recording and its mouth crops are uploaded and separated on a GPU with the live studio’s windows (2 s of context, 0.5 s hop).

  4. Listen, then save

    After separation, switch between Separated and Mixture to compare, and export WAV. The Space stores no audio or video; results stay in your browser.

Need live separation?

ZeroGPU allocates a GPU for each call and counts it against a daily quota, which does not suit a loop that runs every 0.5 s. For live separation and meeting output, run Falcon Studio locally (studio/ in the GitHub repository).

Falcon Studio on GitHub ↗

ZeroGPU gives each Hugging Face account a daily GPU quota, with more quota when signed in. The visible speaker must be synchronized with the microphone audio. Results depend on lighting, framing, background noise, and the model’s training data.