Changelog

• New, early-access beta: Clean my mic. Tick it in the popup and your microphone on Google Meet, the Zoom web client and Discord is cleaned before the other side hears it: background noise, music and nearby talk are turned down, your voice is kept. It uses its own model (mic2); tab filtering is unchanged (vocal12).
• Chrome asks for access to those three sites the first time you turn it on. The access is optional and is given back when you untick the box. Nothing changes for the tabs you filter.
• If anything goes wrong, the call gets your normal microphone. Sites that ask for “original sound” are left alone.
• In our tests it is clearly cleaner than no filter at all, but not yet as clean as the best noise-cancelling tools.
• Not supported: Microsoft Teams in the browser, and calls that ask for camera and microphone in one request (Meet and the Zoom web client often do; use nomus Microphone from the Mac or Linux app there).
• Your microphone is processed on your device and never recorded or sent anywhere. New permission: scripting (to register the call-page scripts). The extension still makes no network requests except the “Rate nomus” link.

• New model, vocal9, for music that is still audible while someone talks or sings. On every test set it leaves 2.0 to 2.6 dB less music under the voice, about 40% less of the music that got through before.
• A small new filter for the low frequencies, where most of that music was. No added delay, and about the same CPU cost.
• Much more of a solo instrument playing on its own is removed: violin, viola, cello, flute, clarinet, oboe, saxophone and brass. The model now learns from separately recorded orchestral parts (the URMP dataset).
• The voice keeps its loudness. The model keeps it slightly quieter, and nomus turns it back up by 2 dB without letting loud passages clip.
• Weaker in one case: clean studio songs keep a little less of the voice than with vocal8. Tenor saxophone in the male-voice range can still pass.
• The new model needs the new engine, and both come in this version. No new permissions. The extension still makes no network requests.

• Microphone (beta): High quality is now the default. It uses DeepFilterNet3, an open-source noise suppressor, and on our first six real Mac recordings it kept your voice closer to the original and sounded cleaner with music and typing around you than Standard (one short take per scene: a first look, not a benchmark). The tradeoff: 20 ms more delay (40 ms in nomus itself instead of 20 ms) and about 1.4% more of one processor core (about 3% instead of under 2%).
• If you used 0.11.0 or 0.11.1, the app moves you to High quality once and says so on the Microphone card. Choose Standard (20 ms, lowest delay, lighter on CPU) any time; nomus keeps your choice from then on. If you had already chosen High quality, nothing changes. (0.11.0 and 0.11.1 saved “Standard” for everyone, so the app cannot tell a deliberate Standard choice from the old default; that is why it switches everyone once and tells you.)
• If the High quality model file is missing or damaged, the microphone keeps running Standard and the card says why, as before.
• The Chrome extension still uses Standard: DeepFilterNet3 cannot run inside the browser’s WebAssembly sandbox yet.
• Packaging: the app no longer contains the user-interface test files and the developer mock that 0.11.1 shipped by mistake (they were never used by the app, but did not belong there), and every build now fails if such files reach the app.
• Your microphone is still processed on your computer and never recorded or sent anywhere. No new permissions and no new network requests. Downloads and updates come from dl.agrohi.com, as for 0.11.0.
• Android stays at 0.9.5. The Chrome extension is not part of this release.

• Fixes the layout of the Microphone card’s Quality control in the app window; no model or format changes.
• The Microphone status line shows “Your microphone” when the system does not report the device’s name.
• Downloads and updates come from dl.agrohi.com, as for 0.11.0.
• Android stays at 0.9.5. The Chrome extension is not part of this release.

• Microphone (beta): a newer default model, mic4b. On our test sets the Standard mode removes about 1.5 dB more music and noise from your voice than 0.10 did, with the same 20 ms of delay. It has not yet been measured on real Mac recordings.
• New, optional: Microphone > Quality > High quality (beta). It uses DeepFilterNet3, an open-source noise suppressor that ships inside the app. It adds 20 ms more delay than Standard (40 ms in nomus itself) and uses about 3% of one processor core instead of about 2%. You can switch while you are in a call; the sound pauses for a moment and comes back. If its model file is missing or damaged, the app keeps Standard and says why. Standard stays the recommended setting.
• Desktop only: High quality is in the macOS and Linux apps, not in the Chrome extension or on Android.
• The app shows the delay as two parts that add up: nomus’s engine and the rest of the path (buffers and the system).
• Downloads and updates for 0.11.0 come from dl.agrohi.com, Agrohi’s download address on Cloudflare R2 (the files grew past the site’s per-file limit). Same operator and domain; the update check and the signature check are unchanged.
• Your microphone is still processed on your computer and never recorded or sent anywhere. No new permissions and no new network requests. DeepFilterNet3 (MIT or Apache 2.0, by Hendrik Schröter et al.) was trained by its authors on public speech-enhancement corpora (the DNS Challenge data, whose noise set includes clips derived from AudioSet and Freesound); nomus does not retrain it.
• Android stays at 0.9.5. The Chrome extension is not part of this release.

• Fixed, Microphone (beta), macOS 14.4 and newer: after installing the driver, nomus could not get macOS to reload its audio, so nomus Microphone did not appear until you restarted the Mac. The app then said both that the driver was in place and that it was not installed. nomus now restarts macOS audio in a way macOS still allows, and checks that it really came back.
• If the driver is on your Mac but macOS has not loaded it, the Microphone card now says “Driver installed but not loaded” and offers Restart audio: macOS asks for your password once and sound pauses for about two seconds. If you had already hit the problem, nomus shows this as soon as you open it. Microphone turns on by itself once the driver loads. If it still does not appear, run sudo killall coreaudiod in Terminal, or restart your Mac.
• Linux: 0.10.2 brings the Linux app to the same version; the fixes above are for macOS only. The filters, models and driver are unchanged. Your microphone is still processed on your computer and never recorded or sent anywhere. No new network requests and no new permissions.

• Microphone (beta), macOS: turning Microphone on now sets up the driver for you. If it is not installed, nomus explains why it needs a small audio driver, macOS asks for your password once, sound pauses for about two seconds, and Microphone turns on by itself. Cancel and nothing changes. If the music filter is on, it pauses too and comes back by itself. When a newer driver comes with nomus, the app offers to update it.
• New and experimental, off unless you turn it on: “Use nomus Microphone as my default input while on”. For apps that have no microphone menu and use the Mac’s default input, such as WhatsApp for Mac. Your previous input comes back when Microphone turns off or nomus quits (the built-in microphone if the old one is unplugged). If you pick another input yourself, nomus leaves it alone. A small helper puts your input back if nomus quits unexpectedly. If your input ever stays on nomus Microphone, pick your mic in System Settings > Sound > Input.
• The installed driver no longer keeps the “downloaded from the internet” flag, which macOS could use to refuse it.
• macOS only: Linux stays at 0.10.0 (no Linux change). The filters and models are unchanged. Your microphone is still processed on your computer and never recorded or sent anywhere. No new network requests and no new permissions. While the default-input switch is on, nomus keeps a tiny local file (its process id and your previous input’s id, no audio).

• New, early-access beta: a Microphone card that cleans what you send in a call. Turn it on, then choose “nomus Microphone” as the microphone in your call app. Background noise, music and other people talking nearby are turned down and the voice at the mic is kept. It runs its own model (mic2) next to the music filter; the music filter itself is unchanged (vocal12).
• In our tests the microphone cleaner is clearly cleaner than no filter at all, but not yet as clean as the best noise-cancelling tools. Your own singing may be turned down too: it is built for speech.
• macOS: the first time, press Install driver on the Microphone card. macOS asks for your administrator password once and reloads its audio for a moment. Uninstall driver removes it again. macOS also asks once for Microphone access. A “Hear yourself” switch plays your cleaned voice to headphones only, never to speakers.
• Linux: nomus Microphone is a PipeWire source that exists while the Microphone switch is on; no driver, no root, and your default devices are not changed.
• Delay: the filter adds 20 ms; from microphone to call app it was about 35 to 40 ms on a Mac and 52 ms on Linux in our measurements. CPU about 2% of one core.
• Your microphone is processed on your computer and never recorded or sent anywhere. No new network requests.
• Not covered: phones (Android and iPhone), Microsoft Teams in a browser, echo from your speakers (wear headphones). The Android app stays at 0.9.5; nothing changes there.
• macOS and Linux update themselves from 0.9.5. A new Mac install still needs the one Terminal line on the download page, because the app is not notarized yet.

• New model, vocal12: acoustic guitar under a voice is cleaned much better. On a guitarist the model never trained on, a guitar playing alone is removed by 28.6 dB (was 22.5 dB), and speech over guitar gains 1.2 to 1.9 dB more.
• It was trained on GuitarSet, 3 hours of acoustic guitar recorded with a microphone (CC BY 4.0). Before this the model had only heard distorted electric guitar.
• Other test sets are unchanged or slightly better: speech gain +0.07 to +0.26 dB, music removed about the same.
• Weaker in some cases: a flute, clarinet or trumpet playing alone is removed 1 to 2 dB less. Under speech they were not worse in our tests.
• Still weak: distorted electric guitar under a voice (about 1 dB below ordinary music), piano and trombone.
• No new permissions. The extension still makes no network requests.

• New model, vocal12, as in Chrome 0.4.3: acoustic guitar under a voice is cleaned much better (a guitar alone is removed by 28.6 dB instead of 22.5 dB on an unseen guitarist; speech over guitar gains 1.2 to 1.9 dB more).
• A flute, clarinet or trumpet playing alone is removed 1 to 2 dB less. Delay and CPU cost are unchanged.
• macOS and Linux update themselves from 0.9.4. Android installs the new APK over the old one.

• Android: the one-time setup no longer needs a computer. On Android 12 or later, nomus uses the phone’s own Wireless debugging: tap Set up, turn Wireless debugging on, and type the pairing code into a notification. nomus then gives itself the permission it needs to silence the original sound and deletes its pairing key. The computer route still works and is still the way for Android 11 and older.
• Android: Start works at once, without the setup. Until the setup is done you hear each app twice (the original and the filtered copy); the setup card says so and is optional.
• Android: the app uses the same fonts and look as the desktop app.
• Android: the app now declares network access, used only by the setup to talk to the phone itself. Nothing is sent off the phone; see the privacy page.
• macOS and Linux: no change to the filter. The version moves with Android so all three match.

• New model, vocal10: 0.6 to 1.1 dB less music under the voice than vocal9 on every test set, and more of a solo instrument playing on its own removed (tenor saxophone 4.6 to 6.8 dB, flute 13.6 to 16.2 dB, violin 13.7 to 16.1 dB).
• It was trained with vibrato and pitch bends added to solo instruments, because a saxophone played like a singer was being taken for a voice. That saxophone is better, and can still be audible.
• The voice keeps its loudness: the model keeps it slightly quieter and nomus turns it back up by 3 dB, without letting loud passages clip.
• Weaker in one case: the hardest clips of clean studio songs keep a little less of the voice than with vocal9.
• Corrects the 0.4.1 notes: the “music alone” figures there were too high, by 0.3 to 5 dB.
• No new permissions. The extension still makes no network requests.

• New model, vocal10, as in Chrome 0.4.2: less music under the voice than vocal9 on every test set, and more of a solo instrument removed. The voice keeps its loudness (a 3 dB boost, behind a limiter).
• The engine reads the larger model formats used by experimental models. Delay and CPU cost are unchanged.
• macOS and Linux update themselves from 0.9.2. Android installs the new APK over the old one.

The first release for macOS and Android, and an update for Linux.

• Less delay on Bluetooth headphones and other outputs that do not run at 48 kHz: nomus adds about 42 ms instead of 73 ms on 44.1 kHz Bluetooth headphones, and about 31 ms instead of 42 ms on Mac speakers. Lip sync in videos holds better.
• The Delay figure shows what nomus adds, and the output’s own delay separately (“~42 ms + 183 ms device”). Bluetooth headphones add their delay with or without nomus, and video players already allow for it.
• If no update is published for your system yet, About says so instead of “Could not check for updates”.
• macOS 14.2 or newer, one download for Apple silicon and Intel. It is signed but not notarized by Apple, so it installs from Terminal (the command is on the download page). Updates install themselves.
• Android 10 or newer: filters apps that allow audio capture, after a one-time setup. No internet permission.
• Model vocal9, as in Chrome 0.4.1 and Linux 0.9.1.

The first release of the Linux app. It filters everything your PC plays, with the same engine and model as the Chrome extension 0.4.1.

• A .deb for x86-64 and ARM64, for Debian-family systems with PipeWire 0.3.49 or newer: Ubuntu 24.04, Debian 12, Linux Mint 22, Pop!_OS 24.04. Install with sudo apt install ./nomus_0.9.1_amd64.deb (or _arm64).
• Filters everything through a PipeWire output called nomus, and puts your own output back when it stops, including after a crash.
• Checks for signed updates at launch and installs them in place; you can turn that off under About.
• Model vocal9: less music under the voice than vocal8, and much more of a solo instrument removed. See Chrome 0.4.1 for the details.

• New model, vocal8. Same size and CPU cost. It removes much more of a solo instrument playing on its own: trumpet 20.4 dB instead of 4.8 dB, flute 10.0 dB instead of 2.9 dB. Every voice test set is as good or better.
• Weaker in one case: singing buried under much louder music in poor audio keeps less of the voice than with vocal6. Tenor saxophone still mostly passes.
• The ON badge clears when the filtered tab closes or navigates away.
• The keyboard shortcut on another tab moves the filter to that tab.
• After ten filtering sessions the popup asks once for a review on the Chrome Web Store. The count stays on your computer.
• No new permissions. The extension still makes no network requests.

The macOS, Linux and Android apps ship the same model, so all four platforms report the same model hash.

• New model, vocal6, trained with more singing. It keeps singing much better: on singers it had not heard, the mean gain rose from +2.6 dB to +5.9 dB. Speech and songs also improved slightly.
• Same network size and CPU cost as before.
• Because the training data now includes GTSinger, the model weights are licensed CC BY-NC-SA 4.0. The credits are in CREDITS.txt.

• Switching modes crossfades instead of clicking.
• Starting a capture can be cancelled, and the icon’s badge only says ON while a tab is really being filtered.
• Internal engine fixes, and support for the newer model file format.
• Same model as 0.3.0.

• New model, vocal4, trained with instrumental tracks from the Free Music Archive. On music it had not heard before, it removes 2.6 to 4.7 dB more of it than the previous model.
• The trade: on songs from the MUSDB test set it gains up to 0.6 dB less than the previous model.