ENGINEER MODE · v25 · LOCAL DSP STACK

Neural Audio Processor — Initializing

Loading ONNX model cache, WhisperHunter AI, 32-stage Deca-Pass pipeline, and spectral controls…

VoiceIsolate

PRO
Engineer Console LOCAL DSP CONSOLE
100%
Stem-Split Download How It Works
--Voice %
--Noise %
-- dBSNR
--Duration
--Sample Rate
--Channels
--RMS
--Peak
--Level
--Voices
--File
IDLEPipeline

PRECISION STUDIO · LOCAL DSP · SINGLE-PASS STFT

Engineer Workstation.

Studio rack for on-device isolation — Process auto-runs analysis & diarization. Export broadcast-ready audio without leaving your device.

Workflow

Ready — upload audio or video to begin

100% local processing · no cloud · no telemetry · no upload
Launch Diagnostics Strictly local processing · WebGPU/WASM inference · one-pass spectral chain
CTX WORKLET GATE DEESS SAB ML ORT NET
Model Cache & Providers

Backend: —

Chain: —

Signal Canvas

Raw immutable · 100% local
Compare A/B switching · Raw immutable
Voice Clarity--
Noise Reduction-- dB
Whisper Retention--
Output dBFS-- dBFS

Selection

Drag on waveform to select · drag rectangle on spectrogram · handles to resize · timestamp shown

Analysis Overlay

Speech Whisper Background Noise Hum Music/bleed

Advanced / Forensic

Deep metrics
SNR: -- dB RMS: -- Peak: -- Level: -- Voices: -- Duration: --
Upload

Drop Audio or Video

MP3 · WAV · FLAC · OGG · OPUS · M4A · AAC · AIFF · WMA · MP4 · MOV · MKV · AVI · WEBM · WMV · TS and more

No file loaded
Library 0 Local OPFS / IndexedDB — never leaves this device

No saved files yet. Choose “Save to library” and upload audio/video.

    Focus on one voice (target enrollment)Local voiceprint · soft gain · no re-ML
    1 · Source Analysis Workspace Confidence —

    Upload above (instant — no decode freeze). Analyze starts local PCM decode and maps voices, whispers, music, horns, barks, and other noise — then shares that map with WhisperHunter so isolation protects speech and suppresses interference. Audio never leaves this device.

    Duration
    Length of the loaded buffer. It is the span Process will cover. Import a file when it stays blank.
    Sample rate
    Rate of the decoded buffer. Processing runs at 48 kHz. A different rate here means the file was decoded at its own rate first.
    Channels
    Mono or stereo on that buffer. Bit depth and codec are not kept. Phase and crosstalk controls apply when the file is stereo.
    Voice %
    Share of energy kept as the signal. If the voice is still buried after Process, raise Voice Iso. That slider does not run the model again.
    Noise %
    Residual against the original, not a separate noise-floor meter. Raise Background Suppress to duck it in playback, or run Process again after noise reduction.
    SNR
    Processed level versus that residual, in dB. A higher number means the kept signal sits further above the residual.
    Peak
    Sample peak in dB. Close to 0 dB means clip risk. Lower the output trim or Limiter Threshold, then Compare.
    Level
    Mean-square level with a −0.691 offset. K-weighting is omitted, so it is not an ITU loudness reading. It stays blank until Process publishes it.
    Voices
    Speaker count from diarization when segments exist. Otherwise 1 when speech energy is present, or 0. Enroll a voice under Focus on one voice if you need one speaker.
    Audit log
    SHA-256 of each processing stage, written locally. Audit Log enables after Process and downloads that history.

    Import a file. Duration, sample rate, and channels come from the decode. Voice, noise, SNR, peak, and level come from Process.

    No recommendation yet
    Run Analyze Full Audio

    Findings

    • Load a file, then analyze.

    Why this preset

    • —

    Source lanes

    Audition
    Detected Sources

    Open-domain sources (speech, hum, residual…) are discovered automatically after Analyze and cached for WhisperHunter. Audition layers below support mute/solo — not a separate process button. 100% local.

    Label Confidence
    Sources appear after Analyze Full Audio (computed automatically).
    Presets & Processing Controls
    2 · WhisperHunter AI

    Best after Analyze Full Audio — collaborates with the workspace above to remove unwanted sound and lift voices/whispers. Uses USM stems when available.

    Noise Reduction / GateInput & pre-cleaning

    Removes unwanted background sound. The gate silences the gaps between words, while spectral NR learns the steady noise (hiss, hum, fans) and subtracts it. Example: clean up a phone interview recorded next to an air conditioner — set NR Amount ~70% and a Gate Threshold around −40 dB.

    EQ (10 bands)Tone & presence

    Ten tone controls from deep bass to bright "air". Boost a band to bring it forward, cut it to push it back. Example: for a thin laptop-mic voice, add ~+2 dB Body and +1.5 dB Presence; for a boomy close mic, cut ~−4 dB Bass.

    DynamicsDynamics & loudness · smoothed

    Evens out volume. The compressor brings loud and soft words closer together; the limiter sets a hard ceiling so nothing clips. Example: a guest who keeps drifting from the mic — Threshold ~−24 dB, Ratio 4:1, Makeup +3 dB for a steady, broadcast-level voice.

    SpectralFilters · de-ess · formants

    Surgical frequency tools: high-/low-pass filters to trim the extremes, a de-esser to tame harsh "s" sounds, tilt for quick brighter/darker, and formant shift for vocal character. Example: a sharp, sibilant narrator — De-ess Amount ~6 dB at 7 kHz, HP Freq 90 Hz to drop rumble.

    AdvancedRoom & stereo

    Restoration and stereo tools: dereverb pulls a voice out of an echoey room, harmonic recovery rebuilds detail lost to noise reduction, and width/phase shape the stereo image. Example: speech recorded in an empty hall — Dereverb ~50% with Rev Decay raised toward 70% for the long tail.

    OutputMaster bus · limiter ceiling

    Final mastering: overall volume trim, the dry/wet blend between original and processed, output stereo width, and export dither. Example: if cleanup sounds too aggressive, pull Dry/Wet to ~75% to mix a little of the natural original back in.

    Separation → IsolationIsolation depth · Live-Mix

    Separation (ML, once per Process) splits voice vs background stems. Isolation sliders below refine that mix in real time — Voice Iso lifts the clean stem, BG Suppress ducks residual noise — without re-running ML. Example: after Process, set Voice Iso ~85% and BG Suppress ~60% to hear a clearer dialogue stem.

    EXTREME / Whisper isolation

    Whisper/forensic isolation refinements that build on separation stems (Process). Whisper Mode Off keeps the ML stem clean; Light–Forensic deepen spectral isolation on the next Process. Live-Mix still uses the separation stems + isolation gains. Example: nightclub whisper at −55 dBFS — Process first, then Whisper Mode Light, Crowd Null ~70%.

    Processing Pipeline
    Ready — drop a file or browse to begin
    S000%
    1. Import decodes the file. Processing runs at 48 kHz.
    2. Process splits voice and background once. One STFT on the fused path.
    3. Live-Mix sliders change playback only. They do not start inference.
    4. Compare switches original and processed without moving the playhead.
    5. Export writes the local file. Drive upload happens only if you ask.
    Duration-- StatusIdle Voice %-- Noise %-- SNR-- Sample Rate-- Channels--
    Ready

    Google Drive is optional. Processing never uses Drive — files leave the device only when you tap Save to Drive.

    0:00 / 0:00
    A Original
    Waveform & Spectrum
    Spectrogram
    Live spectrum rail · scrolling spectrogram · local-only render
    Waveform
    Static inspection before playback · live playhead during transport
    A/B Compare
    Original vs processed waveform overlay for transport A/B
    Level

    Blank until playback. The numbers are a mean-square estimate, not an ITU loudness reading.

    —Window
    —Short
    Clusters
    IN
    OUT
    Input -12.4 dBOutput -10.9 dB
    Aura
    3D Topo
    Swarm
    Liquid