DeepFilterNet3 · WebAssembly · ~3 ms/frame

Clean voice, no noise —
right in the browser

Real-time AI noise suppression for calls, streams and recording. The neural model runs on-device (privacy), with observability, AutoGain and an offline cache.

~60 msend-to-end latency
24 dBnoise suppression
100%on-device
0data sent to server

Why Peito Denoise

More than just "killing noise" — it lets you see and control the pipeline.

🧠

DFN3 neural model

DeepFilterNet3 via onnxruntime-web. Preserves speech, removes noise. Multithreaded wasm inference at ~3 ms/frame — comfortably faster than realtime.

📊

Observability

Live metrics (reduction, latency, VAD, xrun), distortion detectors, input/output spectrograms, WAV dump.

🎚️

AutoGain and high-pass

AutoGain brings quiet speech up to a target level (VAD gate + lookahead limiter), while the input high-pass cuts mic wind, rumble and handling noise before the model.

Low latency

Buffer priming and model warm-up keep end-to-end latency at ~60 ms — no clicks, drops or underruns.

📦

Offline cache

Preloading wasm and models into CacheStorage — instant restart and offline operation.

🔌

Integrations

LiveKit TrackProcessor, simple MediaStreamTrack→Track, bare AudioWorkletNode. Works in Electron/Capacitor.

Pricing

Licenses activate online: the server issues models by key, and revocation takes effect instantly. Bound to domains and a term, managed from the dashboard. Online payment is coming soon — Trial is available right away; for paid plans, email sales@peito.ru.

Trial
Free
14 days · for evaluation
  • 1 domain, 14 days
  • dfn3 neural backend + AutoGain
  • spectral DSP backend
  • Debug panel and metrics
Start free
Enterprise
149,000 ₽/year
for large deployments
  • Everything in Pro
  • Up to 100 domains + wildcard
  • SLA and integration support
Contact

Ready to hear the difference?

Launch the demo bench right now — with your microphone or test files.

Open the demo bench