Skip to audio workspace
VoiceIsolate Pro LOCAL AI STUDIO Download Studio

Voice isolation for recorded audio

Bring the voice forward.

Choose a recording and VoiceIsolate Pro splits it into a clean voice and the background, then lets you remix the two. For interviews, meetings, voice memos and field recordings where speech is hard to hear. Processing happens on this device, so your audio is never uploaded to our servers.

Start Isolating Engineer mode
  • Voice
  • Music
  • Room
  • Air

The picture brings the voice ribbon forward. The mix you control is still voice and background.

Simple on the surface.
Strict underneath.

One forward STFT. In-place spectral work. One inverse STFT. Audio is decoded locally; batch ML never runs from a slider.

  1. 01

    Input

    Upload-only file pick. No microphone.

  2. 02

    Decode / resample

    Local decode to 48 kHz buffers.

  3. 03

    Stem separation

    Batch ONNX in a worker (BS-RNN default).

  4. 04

    Spectral processing

    Fused masks on one STFT representation.

  5. 05

    Reconstruction

    One inverse STFT. No extra phase passes.

  6. 06

    Live-Mix

    Zero-ML playback graph: gains, EQ, gate, de-esser.

  7. 07

    Export

    Local encode. Optional user-initiated Drive save.

Batch ML

Explicit Process / Separate. ONNX in MLWorker. Sliders do not retrigger it.

Live-Mix

AudioParam + Gate/De-Esser worklets on already-separated stems. Immediate, reversible.

Privacy

No cloud inference. Optional Drive I/O is user-initiated file transfer only.

Built for

Meetings

Isolate voices from background noise, hum, and cross-talk. Automatic speaker detection with confidence.

Field Recording

Recover whispers and faint speech in noisy environments. Works offline in the field once it has loaded.

Investigation

Forensic inspection with spectrogram time-frequency selection and Removed (Delta) audition.

Content Creation

Clean voice for podcasts and video. Preserve raw as immutable, process non-destructively.

02

Separate & mix

Audio or video · decoded at 48 kHz on your device. Nothing is uploaded to a server. Analysis starts automatically after import.

Local session · audio stays here

Import

Drop audio or video here

WAV · MP3 · M4A · FLAC · OGG · MP4 · MOV and more

Profile

Import starts automatic local analysis — noise floor, VAD, whisper candidates, speakers. Profiles change processing behavior; Whisper Boost appears as a contextual action when relevant.

Checking local processing support…

Idle — choose a file to begin. Audio stays on device.

Idle — choose a file to begin Inference: local ONNX

Signal workspace · select directly on waveform or spectrogram

Stems

Idle
Voice
100%
Background
100%
0:00 / 0:00
100%

Compare · Raw / Processed / Removed

Audition what processing removed so accidental speech destruction is detectable. Raw is immutable.

Live output spectrum — reacts to Live-Mix sliders in real time
Source confidence appears after on-device analysis. No placeholder measurements.
Advanced mix · EQ · dynamics · gate
0 dB
0 dB
0 dB
0 dB
0 dB
20 Hz
20000 Hz
0 dB
1:1
20 ms
250 ms
0 dB
0 dB
100%
-45 dB
0 dB
5 ms
100 ms
6000 Hz
0%
Advanced / Engineer / Forensic inspector

Analyze · Automatic

Source separation, VAD, whisper detection, noise classification, mask generation happen automatically behind the scenes as part of local analysis. Overlays: Speech, Whisper, Background Noise, Hum, Music / bleed, Candidate secondary speaker.

No analysis yet — import audio to see automatic regions with confidence.

Need per-band EQ, dynamics, gate, per-speaker lanes or forensic views? Open the Engineer Console.

03

Engineer Console

The full precision workstation: signal-first canvas, 67 controls, source lanes and forensic inspection, all running locally.

Engineer Console: signal canvas, source lanes and DSP rack Open the Engineer Console

Open the console in a mode

Quick

Guided cleanup

One primary action: import, automatic analysis, recommended enhancement, A/B, export.

Open Quick
Meeting

Meeting focus

Optimized for multi-speaker isolation and background suppression.

Open Meeting
Studio

Precision workstation

Default professional shell with signal-first canvas and full local DSP.

Open Studio
Forensic

Signal inspection

Spectrogram-first workspace with time-frequency selection and advanced inspector.

Open Forensic

Trust · Privacy

Local processing. No cloud upload, no telemetry.

100% local processing with WebGPU/WASM. No cloud upload required, no telemetry for audio processing, and you keep control of the original audio. Raw source is immutable; processed and removed are derived views. No fabricated customer counts or quality percentages.

On-device Private Zero Cloud Offline after first use

Settings · Privacy & local processing

Models run in a Web Worker on your device (WebGPU when available, WASM fallback). No microphone access, no audio upload to our servers, no telemetry.

Backend: detecting…

Start isolating Get Android / Windows