Introduction

Hush AI shipped as an open desktop app this window: alshell7/hush-ai releases v1.0.0 → v1.2.0 on 2026-11-03, with model files mirrored at alshell7/hush-ai on Hugging Face. It is a local noise-cancellation tool for live calls and offline Studio cleanup — not a new TTS/ASR foundation model, and not Ema’s unrelated “Hush” PII redaction model.

Site + docs: alshell7.github.io/hush-ai.

What Shipped

  • Vendor / product: Hush AI — alshell7
  • Modality: on-device audio tool / real-time speech enhancement (virtual mic + Studio)
  • Open vs closed: app + engine code Apache-2.0; DeepFilterNet 3 weights MIT/Apache; NVIDIA Real-time RE-USE weights under NVIDIA One-Way Noncommercial (unchanged upstream)
  • Runnable today: download the latest GitHub Release for Windows / macOS / Linux, or build from source; models auto-fetch (or pull the HF mirror)
  • Concrete fact: default DeepFilterNet 3 path uses 10 ms frames (~60 ms before device buffers); RE-USE “Light” quality on Apple M1 (2 threads) clocks p99 11.7 ms per 20 ms frame, with CPU output matching NVIDIA’s GPU reference at 113–118 dB waveform SNR on short speech in CI

Hush installs a Hush AI Microphone virtual input for Zoom, Meet, Teams, Slack, Discord, Webex, OBS, plus optional Hush AI Speaker cleanup for what you hear. Studio cleans podcasts and voice notes on-disk. Live Translate is a separate Gemini-powered preview; noise cancellation itself stays offline.

Run It

huggingface-cli download alshell7/hush-ai --local-dir hush-ai-models

Desktop builds: GitHub Releases latest — then choose Hush AI Microphone in your call app.

Rust makers can drive RE-USE through the hush-core crate bundled in the repo (see the HF card for the Denoiser::with_config snippet). Switch engines in the app UI: DeepFilterNet 3 (default, commercial-friendly) vs NVIDIA RE-USE (Light / Balanced / Studio quality tiers).

Notes for makers

  • This is a day-0 open Krisp/Adobe-Podcast-style tool, not new foundation weights — value is the runnable desktop surface + dual-engine packaging with named latency numbers
  • RE-USE remains noncommercial upstream; DeepFilterNet 3 is the path if you need commercial-friendly licensing
  • Calendar join + Live Translate are conveniences; keep expectations clear that NC never leaves the machine while Translate uses Gemini when enabled
  • Soft-holds unchanged this cycle for gated/401 TTS (Ghana ZipVoice, Moore, finanfa Fon TTS, Underdog Woof HF 401) — Hush clears the write filter as an audio tool with a public release artifact inside ~48h
  • No Gen CTA: desktop NC does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

For makers who want free, inspectable, on-device call cleanup today — with DeepFilterNet speed or RE-USE quality on tap — Hush AI v1.2.0 is the open desktop drop to try before another closed NC subscription.

—Anabel ♡