Free alternative to Krisp & Adobe Podcast

Silence the noise.
Keep your voice.

Hush AI removes keyboards, fans, barking dogs, traffic and café chatter from your microphone while you talk, and cleans up recordings, on your own computer. Every app gets a clean microphone called Hush AI Microphone, and with Live Translate it can even speak for you in another language.

Download Hush AI View on GitHub

  • Free, no limits
  • Open source
  • No account
  • Windows, macOS, Linux
Keyboard
Fan hum
Traffic
Your voice
Works with
  • Zoom
  • Google Meet
  • Microsoft Teams
  • Slack
  • Discord
  • Webex
  • OBS Studio
  • and more
Listen

Hear the difference

Pick a scene and press play. Switch between the original and Hush while it plays: both run in sync, so you hear the same word, with and without the noise.

With Hush · NVIDIA RE-USE
0:00 / 0:06 · NVIDIA's sample recording
Hush engine
  • PESQ–
    sound quality
     
  • STOI–
    clarity
     
  • SI-SDR–
    noise removed
     

NVIDIA's sample recording, before and after Hush (NVIDIA RE-USE). Sample © NVIDIA (license). How we measure

How Hush was built

From a research paper to your call

Five layers, from GPU research code to a microphone every app can use. Scroll through them.

GPURE-USEDFN3PyTorch110–20 ms framesWindows · 100+ dBmacOS · 100+ dBLinux · 100+ dB2ringaudio threadworkeroutput3Core AudioVB-CABLEPulseAudio45
  1. 1

    Two research models

    NVIDIA Real-time RE-USE, a Mamba speech model from NVIDIA research, written in PyTorch for GPUs. And DeepFilterNet 3, an open model built for real-time noise suppression.

    • PyTorch
    • GPU
    • Mamba
    • DeepFilterNet 3
    How the two compare
  2. 2

    Ported to plain Rust, on the CPU

    RE-USE was rewritten in plain Rust with SIMD; DeepFilterNet 3 runs through its own Rust library, libDF. Both run in real time on the CPU, in 10–20 ms frames. On every change, CI checks RE-USE against NVIDIA's GPU output on Windows, macOS and Linux: the waveforms match to more than 100 dB.

    • Rust
    • SIMD
    • No Python
    • No GPU
    • 100+ dB in CI
    Accuracy
  3. 3

    A real-time engine

    The audio thread never waits. Lock-free ring buffers carry your voice from the microphone to a worker thread that runs the model, and on to the output. A slow frame can never glitch the device.

    • Lock-free queues
    • Inference on a worker
    • Clock-drift control
    Architecture
  4. 4

    The app and a virtual microphone

    A small Tauri app with a tray widget, and a virtual microphone for each system: Hush's own Core Audio plug-in on macOS, VB-CABLE on Windows (until Hush's own driver is signed), and a PulseAudio virtual source on Linux.

    • Tauri
    • Core Audio plug-in
    • VB-CABLE
    • PulseAudio
    Virtual microphone
  5. 5

    Every meeting app

    Apps see Hush AI Microphone like any other microphone. Zoom, Google Meet, Teams, Slack, Discord, Webex, OBS: choose it once, and they hear only you.

    • Zoom
    • Meet
    • Teams
    • Slack
    • Discord
    • Webex
    • OBS
    Performance
Compare

The free, open-source alternative

Hush AI does the job of Krisp for live calls and of Adobe Podcast Enhance Speech for recordings. Free, with no account, and your audio stays on your computer.

Hush AIKrispAdobe Podcast Enhance Speech
PriceFree, no limits7-day free trial, then paid plans (Core $8 per user a month, billed yearly)Free with daily limits; more with a paid Adobe plan
AccountNoneSign-up requiredAdobe account
Where your audio is processedYour computerYour computer (noise cancellation)Adobe's servers: you upload the file
Open sourceYes, Apache-2.0NoNo
Live calls in any appYesYesNo
Clean up recordingsYes, on your computerNot found on its pagesYes, in the browser

From the vendors' public pages (krisp.ai/pricing, podcast.adobe.com), November 2026; – means we did not find it there. We have not measured Krisp's or Adobe's audio quality. Hush's optional Live Translate (Preview) sends your voice to Google while it is on. Krisp and Adobe are trademarks of their owners, who do not endorse Hush AI.

Features

Quiet, fast and yours

Two proven speech models in native code, right on your CPU. No GPU, no cloud, no subscription. How it works

  • On your computer, and private

    Your voice is cleaned on your own CPU. There is no account and no telemetry. Hush goes online only for the one-time model download, update checks, your calendar if you connect one, and Google while you use Live Translate, the one feature that is not on-device.

    • No account
    • No telemetry
    • Noise cancellation works offline

  • Real time

    About 60 ms from your microphone to your call. Each model needs about 40 ms on its own; Hush works in 10–20 ms frames.

  • Light on the CPU

    DeepFilterNet 3, the default engine, uses about 7% of one CPU core. Your fans stay quiet and your battery lasts.

  • Works with every app

    Hush adds a microphone called Hush AI Microphone. Choose it once in Zoom, Meet, Teams, Slack, Discord, Webex or OBS.

  • The notch

    Hush lives at the top of your screen, wakes up in calls and opens when you point at it. Prefer a tray widget? One setting. Both can hide from screen sharing.

  • Automatic updates

    New versions download quietly. Each one is signed, and Hush checks the signature before it installs.

  • Windows, macOS and Linux

    Windows on x64 and ARM64. macOS 11 or later on Apple silicon and Intel. Linux as an AppImage, .deb or .rpm.

    • Windows x64 · ARM64
    • macOS 11+ · Apple silicon · Intel
    • Linux · AppImage · deb · rpm
Works in every app

One microphone. Every app.

Hush adds a microphone called Hush AI Microphone. Choose it once in your meeting, chat or streaming app, and the app hears only your voice. Point at an app to see where.

App names and logos are trademarks of their owners, who do not endorse Hush AI. They are shown only to say which apps work with Hush AI Microphone.

New · Enhance a recording

Clean up a recording. Keep it here.

Drag an audio file into Hush, or record yourself, and Hush removes the noise on your own computer. The same job as Adobe Podcast Enhance Speech, with no upload, no account and no daily limit.

  1. 1Drop a file, or record yourselfAn interview, a lecture, a voice note or a podcast take.
  2. 2Hush cleans it on your computerNothing is uploaded. No account, no daily limit.
  3. 3Compare before and afterSwitch between the original and the clean version while it plays.
  4. 4Save a clean WAVReady for your editor, your podcast or your message.
  • Opens
  • WAV
  • MP3
  • M4A / AAC
  • FLAC
  • OGG
New · Live Translate (Preview)

Talk in your language. Understand theirs.

Live Translate works both ways: the call hears you in their language, in a natural voice, and you read (or hear) them in yours. It works in every app through Hush AI Microphone and Hush AI Speaker, with Google's Gemini Live Translate model.

The one Hush feature that is not on your computer

  • While Live Translate is on, the voices you chose go to Google to be translated. Hush itself has no server.
  • Noise cancellation, recordings and everything else stay on your computer.
  • The translation follows about a sentence behind you: Gemini translates phrase by phrase.
  • It needs an internet connection.
  1. 1Bring your own Gemini API keyHush opens Google AI Studio's key page, notices the key you copy, checks it with Google and keeps it in Hush's settings folder. Google has a free tier.
  2. 2Pick the languages“I speak” and “They speak”, from more than 70 languages, with a swap button. Both ways, or just your voice, or just theirs.
  3. 3Turn it on in your callHush asks once. They hear you in their language through Hush AI Microphone; you read them in yours, or hear a voice-over through Hush AI Speaker. Nothing changes in Zoom, Meet or Teams.

Switch it during a call

  • The notch at the top of your screen, with both sides' captions
  • The Talk tab, where every call is kept as a conversation you can search and export
  • “Translate my voice” and “Translate their voice” in the tray menu
  • A shortcut:Ctrl+Alt+Ton Windows and Linux,⌘⌥Ton macOS

Live captions show what you said and what they hear. Languages use BCP-47 codes: en, es, fr, de, hi, ar, zh-Hans, ja, pt-BR and more.

New · Google Calendar

Ready before the meeting starts

Connect Google Calendar without signing in. Hush shows your next meetings with a Join button, and turns noise cancellation on just before a call.

Sync settings

Try them: this is how they look in Hush.

Read the calendar every
Turn Hush on before meetings with a call link

Private by design. The secret address stays on your computer. Hush reads it straight from Google: there is no Hush server and no Google sign-in.

The app

One switch. That's it.

Turn Hush on, pick your microphone, and get back to your call.

Noise cancellation On
Your voice, without the noise

Apps hear you through Hush AI Microphone.

MicrophoneMacBook Pro Microphone▾
Your microphone
Hush AI Microphone
Hush AI
Noise cancellation
On · apps hear Hush AI Microphone
MacBook Pro Microphone▾

Always one click away

Hush lives in your system tray on Windows and Linux, and in the menu bar on macOS. The small widget does the everyday things without opening the window.

  • Turn noise cancellation on or off
  • Switch microphones
  • Glance at your levels
Benchmarks

Two engines. Measured.

On the VoiceBank-DEMAND test set (824 recordings, scored at 16 kHz), the best score on every quality metric belongs to one of Hush's engines.

Perceived quality · higher is better

How close the result sounds to the clean studio recording, on a scale of 1 to 4.5.

  1. No suppression
    1.97
  2. RNNoise
    2.11
  3. HushNVIDIA RE-USE Light
    2.61
  4. HushNVIDIA RE-USE
    2.77
  5. HushDeepFilterNet 3
    3.10 (best)
1.004.50

One minute of audio, one CPU core

How long each engine needs to clean 60 seconds of speech on one thread. Played 10× faster.

  • RNNoise3.0 s
  • DeepFilterNet 34.2 s

In the app, Hush uses more threads when one is not enough. Times from the CPU column: 60 s × CPU per audio second.

Why two engines?

DeepFilterNet 3 is light, scores highest on PESQ, STOI and SI-SDR, and is free for any use, so it is the default. NVIDIA RE-USE, with 8 layers, scores the most natural sound (DNSMOS), but it needs much more CPU. You choose in Settings.

The default

DeepFilterNet 3

  • About 7% of one CPU core
  • Highest PESQ, STOI and SI-SDR here
  • Full band, 48 kHz, on one thread
  • Free for any use, also at work

License: MIT or Apache-2.0. Free for any use.

The most natural sound

NVIDIA Real-time RE-USE

  • Best DNSMOS (natural sound) with 8 layers
  • Light, Balanced and Studio quality in the app
  • Wants a faster CPU
  • Optional: download it in Settings

License: NVIDIA RE-USE model license

  • DeepFilterNet 3 numbers come from the official deep-filter 0.5.6 binary. Hush runs the same libDF code and model.
  • One dataset, objective metrics only, no listening test. The test speech and noise are close to DeepFilterNet 3's training data.
  • CPU is approximate: one process at a time on an 8-core x64 laptop that was also doing other work.
  • Live load in the app (debug build, Windows laptop, one thread): DeepFilterNet 3 about 2.3% of real time.
The method, every metric and the caveats
Roadmap

Noise cancellation is just the start

Noise cancellation is here, and voice translation is in preview. The next ones are open for ideas and for help.

  • Noise cancellation Available
  • Voice translation Preview
  • Agent assistComing soon
  • Speech analyticsComing soon
  • Voice securityComing soon

Want to help build the next one? Ideas, issues and pull requests are all welcome.

Contribute
FAQ

Questions, answered

  • Yes. Hush AI is free and open source, with no account, no trial, no subscription and no minute limits. The code is on GitHub.

Give your calls some quiet

Free for Windows, macOS and Linux. No account, no limits.

Download Hush AIAll releases

Windows

macOS

Linux

Builds aren't code-signed yet, so your computer asks once before opening Hush. Here's what to click. On Windows, also install the free VB-CABLE for now.