Phonon for Windows beta 4 and Linux preview

Phonon for Windows, beta 4, and Linux preview

A more accurate speech recogniser. Everything else is as in beta 3.

What's new

• Phonon's own speech model (N8avg). Parakeet TDT 0.6B v2 with Phonon's N8avg fine-tune merged in. It was trained only on public, openly licensed speech, with no personal recordings. On 578 real dictation clips it makes fewer errors than the stock model (word error rate 0.49 points lower, 95% CI 0.28 to 0.71). On the public LibriSpeech recordings used to test long dictations, raw word error rate is 1.83% on the 410 s recording (unchanged) and 1.35% on the 900 s one (was 1.69%). After correction it was 1.83% / 1.35% on Windows and 2.10% / 1.48% on Ubuntu; correction varies by a few words from run to run.
• The same files as the Android keyboard. Both download from Infatoshi/phonon-asr-parakeet-tdt-0.6b-v2-onnx, pinned to commit 76fbda42 and checked against its sha256 before use. The encoder's weights now sit in a separate file that is read from disk as needed.
• Updating downloads the new speech model once (about 630 MB) and then deletes the old one. The correction model is unchanged.

Windows (beta 4)

Run Phonon-Setup-1.0.0.exe. The file still carries version 1.0.0. Tested by installing over an earlier build in a Windows 11 VM that held the stock speech model: the update fetched the new model, removed the old one, and transcribed and corrected both long LibriSpeech recordings in full.

Not code signed: SmartScreen shows "Windows protected your PC" the first time. Choose More info, then Run anyway.

Linux (preview)

sudo apt install ./phonon_1.0.1-beta4_amd64.deb

The package version is 1.0.1~beta4, so it upgrades beta 3. Tested on Ubuntu 26.04 by installing over beta 3. Requirements and setup are unchanged: Ubuntu 24.04 or later on x64, GNOME or KDE Plasma, Wayland or X11, hold Right Ctrl to dictate.

Known issues

• English only.
• Tag questions and leading answer words ("..., right?", "Yeah, ...") are often removed as filler.
• Windows: balloons do not stay in the notification center.

Checksums (sha256)

42a7b04a85a49f12d6e7fbab542c1ee96400a828d47c98265f6293fdc52e4b51  Phonon-Setup-1.0.0.exe
ff05553fce9bc2a4d28393bcc51fb20ad9688ff039172b986cb0f906b9608a59  phonon_1.0.1-beta4_amd64.deb

Speech model files (sha256)

5c0a15da8929acd0ddd559471a201bd8ede3bbe2a7f2df6f606d68b8b73f52bb  encoder.int8.onnx
f91c1c8d1c8269c9d21d4d23e7590b924243ce6036a1ff787e0e1beb515dabca  encoder.int8.onnx.data
b6bb64963457237b900e496ee9994b59294526439fbcc1fecf705b31a15c6b4e  decoder.int8.onnx
7946164367946e7f9f29a122407c3252b680dbae9a51343eb2488d057c3c43d2  joiner.int8.onnx
ec182b70dd42113aff6c5372c75cac58c952443eb22322f57bbd7f53977d497d  tokens.txt

License

GPL-3.0. The speech model is CC-BY-SA-4.0 (NVIDIA Parakeet, CC-BY-4.0, adapted with EdAcc, CC-BY-SA-4.0, among other public sets); the refiner weights are under the LFM Open License v1.0; the runtimes are sherpa-onnx (Apache-2.0) and llama.cpp (MIT).

Phonon 1.0.1

Fixes from the launch feedback, a few features people asked for, and a correction model that puts the question mark back.

What's new

• Questions end with "?". Parakeet often ends a spoken question with a period, and the old model kept it. The new correction model ends a question with "?" in every writing mode, casual included, without adding one to statements: on a 40-question probe in minimal mode it went from 19 to 40 right (casual from 0 to 40), with no false "?" on 40 look-alike statements in any mode.
• Answer words and confirmation tags stay. A leading "Yeah," or "No," and a closing "..., right?" are kept instead of being cut as filler.
• Better text overall. On 2,522 real dictations, a local judge preferred the new model's text over 1.0's 575 times to 248 (1,200 came out the same), and it kept the meaning better 156 to 98. Details on the model card.
• Tap to start and stop. Settings, under Shortcut, "Tap to start and stop": one tap of Globe (fn) or Right Option starts dictation, the next press stops it. Holding still works as push-to-talk. Off by default. (#34)
• Lists and emails get line breaks. Dictated steps become a numbered list with one item per line in every mode but casual, and in email a dictated greeting and sign-off go on their own lines. Casual chat stays one paragraph.
• Writing mode per app. Settings > Advanced lets you pin minimal, casual, formal or prompt for any app, overriding the automatic pick.
• Your own model, if you want one. "Your own model endpoint" in Settings can send the transcript to any OpenAI-compatible endpoint you choose, with the key kept in your Keychain. Off by default; there is no Phonon server. Phonon uses the endpoint's text only if it arrives within 6 seconds and passes the same safety check as the local model, otherwise the local result. Speech recognition always stays on your Mac. (#24)
• Other audio gets out of the way. While you hold the key, music or a video playing through your speakers fades down (or mutes, your choice in Settings > Dictation) so it does not end up in the transcript, then fades back.
• New microphone? Phonon asks once whether to use it first.
• Import from other dictation apps. Bring your dictionary, replacements, snippets, history and shortcut over from Wispr Flow, Aqua Voice, superwhisper, VoiceInk, MacWhisper or FluidVoice. Dictations into private apps stay out.
• Whispers count. The silence check lets more softly spoken dictations through (it used to drop close to half of whispered clips).

Fixes

• Left Option no longer starts a Right Option dictation. Thanks @oazabir (#29, fixes #23).
• Control-Space is only reserved when your shortcut uses it, so it switches input sources again in the default setup. Thanks @jmmarotta (#32).
• Globe (fn) is more reliable: a press after a lost key release now starts dictation instead of being ignored, and two quick taps that recorded nothing show a hint to hold the key. If macOS is set to do something else when you press Globe, Phonon now tells you. (#21)
• Every card above the capsule has a close button. (#22)
• Phonon tells you when Input Monitoring is off, on every page and in the menu bar, instead of silently ignoring the hold key.
• The MLX fallback's memory no longer keeps growing with every dictation. Thanks @jmmarotta (#33, fixes #27).

Updating

• Updating downloads the new correction model, about 2.1 GB (its GGUF and its MLX fallback). Parakeet is unchanged.

Known issues

• In Slack, Teams and Google Chat (formal mode), a dictated greeting such as "Hi Sam," can land on its own line.
• "Now what?"-style questions are often trimmed.
• About half of self-corrections are still missed.
• Tested on M4 and M4 Max. M1 to M3 have not been tested on hardware yet; if the Metal engine fails its startup self-test there, Phonon uses MLX as before.
• English only.

Install

brew install --cask infatoshi/tap/phonon

or download Phonon.dmg below (signed and notarized). Existing installs see the update within a day, or run brew upgrade --cask phonon.

License

Phonon stays GPL-3.0. The refiner weights are under the LFM Open License v1.0. Parakeet is CC-BY-4.0 (NVIDIA). The engine links llama.cpp (MIT, commit 39febd2 plus the patch in third_party/llama.cpp).

Phonon for Windows beta 3 and Linux preview

Phonon for Windows, beta 3, and Linux preview

Long dictations now work. Built from the same source as Phonon 1.0.1 for Mac, plus the fixes below. Correction still uses the refiner from the 1.0 release, as in beta 2.

What's fixed

• Dictations over 6 min 40 s no longer crash Phonon. The speech model accepts at most 400 seconds of audio in one pass. A longer recording made it throw, Phonon closed, and the recording was lost. Recognition now runs in passes of at most 60 seconds, each cut at the quietest moment near the end of its window. Shorter passes also drop fewer words. On a 15-minute LibriSpeech recording, word error rate was 1.35% with 60-second passes vs 2.83% with 6-minute passes. If a pass still fails, Phonon reports an error instead of closing.
• Long transcripts are corrected in chunks. Correction now works through a long transcript in groups of whole sentences that fit the model's context. Before, a 7-minute dictation lost the end of its text, and a 15-minute one was rejected outright. If correction fails, Phonon types the raw transcript instead of nothing.
• Long recordings are saved first (Windows). A recording over a minute is written to the history folder before transcription starts, so the audio survives if anything goes wrong.

Windows (beta 3)

Run Phonon-Setup-1.0.0.exe. The file still carries version 1.0.0.

Not code signed: SmartScreen shows "Windows protected your PC" the first time. Choose More info, then Run anyway.

Linux (preview)

sudo apt install ./phonon_1.0.1-beta3_amd64.deb

The package version is 1.0.1~beta3, so it upgrades beta 2's 1.0.0 package. Tested on Ubuntu 26.04 by installing over beta 2: public LibriSpeech recordings of 410 s and 900 s were transcribed and corrected in full (word error rate 1.64% and 1.61% after correction), where beta 2 crashed. Requirements and setup are unchanged from beta 2: Ubuntu 24.04 or later on x64, GNOME or KDE Plasma, Wayland or X11, hold Right Ctrl to dictate.

Known issues

• English only.
• Tag questions and leading answer words ("..., right?", "Yeah, ...") are often removed as filler.
• Windows: balloons do not stay in the notification center.

Checksums (sha256)

e896213818fadf5ad41106e268bb9c549f1d52fd83c92fa834c9d281a54e0c25  Phonon-Setup-1.0.0.exe
212324ad6b33af4addc790145e4a35d16a91491c875651b835d579ad941c3b0d  phonon_1.0.1-beta3_amd64.deb

License

GPL-3.0. The refiner weights are under the LFM Open License v1.0; Parakeet is CC-BY-4.0 (NVIDIA); the runtimes are sherpa-onnx (Apache-2.0) and llama.cpp (MIT).

Phonon for Windows beta 2 and Linux preview

Phonon for Windows, beta 2, and Linux preview

Hold a key, talk, let go, and the text appears where you are typing. Speech recognition (Parakeet) and correction (Phonon's refiner, the same model as the Mac app's 1.0 release) run on your computer. First start downloads about 1.6 GB; after that everything is offline.

Built from the same source as Phonon 1.0.1 for Mac. The files carry version 1.0.0.

Windows (beta 2)

• A real installer. Double-click Phonon-Setup-1.0.0.exe: it installs for your user only (no administrator needed), adds Phonon to the Start menu and to sign-in, and starts it. If the Microsoft Visual C++ runtime is missing, it installs it and Windows asks once for permission. Uninstall from Settings > Apps.
• The refiner replaces Gemma: a smaller download (1.6 GB instead of 3.8) and the same correction model as the Mac.
• Correction runs on your graphics card through Vulkan when one is fast enough, otherwise on the CPU. On an RTX 3090 a short dictation's correction takes about 65 ms.
• Pick the dictation key from the tray menu: Right Ctrl (default), Right Alt, Ctrl+Space, Right Shift, Caps Lock or F13. Laptops with a Copilot key instead of Right Ctrl switch to Right Alt by themselves the first time you press it.
• Phonon warns you when the window you are typing into runs as administrator, because Windows does not let it type there.
• A busy computer no longer stalls dictation: with every core loaded (a compile, a render), dictations landed in 1.5 to 4 seconds in our tests, where earlier builds could take minutes.
• Fixes: pressing the key while Phonon is still setting up now says so instead of replaying later; memory use is about 0.8 GB lower; an empty recording no longer crashes Phonon; the correction server closes with Phonon.

Not code signed: SmartScreen shows "Windows protected your PC" the first time. Choose More info, then Run anyway.

Linux (preview)

sudo apt install ./phonon_1.0.0_amd64.deb

• Ubuntu 24.04 or later and Debian-family systems on x64, GNOME or KDE Plasma, Wayland or X11.
• Hold Right Ctrl to dictate. Phonon reads the key from the kernel, so it works under Wayland, and types by pasting with a virtual Ctrl+V, putting your clipboard back afterwards. In a terminal the text stays on the clipboard (terminals paste with Ctrl+Shift+V).
• The package installs a udev rule so your user can read the keyboard and create a virtual one; no group change or logout. It starts at sign-in.
• Correction runs on the CPU.

Known issues

• English only.
• Tag questions and leading answer words ("..., right?", "Yeah, ...") are often removed as filler.
• Windows: balloons do not stay in the notification center.

License

GPL-3.0. The refiner weights are under the LFM Open License v1.0; Parakeet is CC-BY-4.0 (NVIDIA); the runtimes are sherpa-onnx (Apache-2.0) and llama.cpp (MIT).

Phonon 1.0

Phonon 1.0 (macOS)

Phonon writes what you meant to type, in the style of the app you are typing into, and now shows it while you talk.

What's new

• Live text. While you hold the key, a card above the capsule shows the corrected text so far, with the words just heard after it. It changes only while you are speaking, and words more than about a dozen back stay put until you let go. Releasing the key pastes the final text. On by default; turn it off in Settings. It needs streaming correction, which runs on Macs with 16 GB of memory or more.
• Long dictations stay fast. Past about 200 words, Phonon corrects the older sentences once and keeps them, and only the newest words are corrected again at key up. A 6-minute dictation replayed through the engine came back 0.2 s after key up instead of 7.6 s, with fewer filler words left in.
• A new main window: Home (your average speed and your last dictations, each with what you said and what Phonon wrote), Notes (meeting recordings), Insights, Dictionary, Macros, and a simpler Settings.
• Nothing gets lost. If nothing can take the text (the desktop, a window without a text field), Phonon does not paste into the void: it keeps the text on a card with a big Copy button, and it is on your clipboard too. The same card tells you when the mic sends only silence, when nothing was heard, and when macOS has not allowed Phonon to paste.
• Clipboard choice. Phonon can leave the dictation on the clipboard (as before) or put back what you had copied. First run asks; Settings > Dictation changes it.
• A new correction model. It removes filler and false starts. When you correct yourself mid-sentence ("tuesday, actually no, thursday"), it keeps only the final version about half the time (0.48 to 0.58 on a 96-item test, depending on the mode); 0.2.3 kept both versions every time. It is fine-tuned from LFM3-1.7B with no personal dictations; the model card on Hugging Face lists the training data and every number.
• Dictionary terms hold. A one-word term from your dictionary that you said is kept when the model would swap it for a different term (it wrote "MTK" for "MTP" in a long dictation).
• Writing modes follow the app:
  • casual in personal chat (Discord, Messages, WhatsApp): lowercase, no final period;
  • formal in email and work chat (Slack, Teams, Google Chat); in email a dictated greeting and sign-off go on their own lines;
  • prompt in AI assistants, terminals and code editors;
  • minimal everywhere else.
    In a browser the site decides, so Gmail is email and claude.ai is prompt.
• Your dictionary now steers speech recognition itself. Parakeet is biased toward your terms while it decodes, so jargon comes out spelled your way before the correction model sees it. Ordinary English words in the dictionary are left out, because boosting them could only insert them where they were not said.
• Phonon tells you when a newer version is out. Once a day it asks GitHub for the latest release; the request carries nothing about you or your dictation. It shows one alert per version and never interrupts a dictation. Turn it off in Settings > System. This release is the first with the check, so it starts working with the next release.
• Quiet dictations no longer vanish. A loudness check used to throw away clips it judged silent, and it caught some softly spoken sentences. Now speech recognition gets the final say: three words or more always go through.
• Recordings keep the whole dictation (up to 20 minutes); before, only the last 2 minutes were saved.
• Updating downloads the new correction model, about 2.1 GB (its GGUF and its MLX fallback). Parakeet is unchanged.

Known issues

• About half of self-corrections are still missed. When the model misses one, it may leave the sentence as you said it, filler included.
• Words the speech model mishears as a different real word (for example "Quinn" for "Qwen") are not always fixed; add a spoken form to the dictionary entry.
• Connecting or disconnecting Bluetooth headphones (for example AirPods) while Phonon is idle can crash it. Reopen Phonon.
• Tested on M4 and M4 Max. M1 to M3 have not been tested on hardware yet; if the Metal engine fails its startup self-test there, Phonon uses MLX as before.
• English only.

Install

brew install --cask infatoshi/phonon/phonon

or download Phonon.dmg below (signed and notarized).

License

Phonon stays GPL-3.0. The refiner weights are under the LFM Open License v1.0. Parakeet is CC-BY-4.0 (NVIDIA). The engine links llama.cpp (MIT, commit 39febd2 plus the patch in third_party/llama.cpp).

Phonon 0.2.3

Phonon 0.2.3 (macOS)

Fast speech recognition on every Apple Silicon GPU, not just the big ones.

What changed

• Speech recognition's Metal decoder no longer needs every part of the GPU to run at once. On Macs with smaller GPUs (a base M4 MacBook Air, for example), 0.2.2's decoder timed out and Phonon fell back to its slower MLX path on most dictations. The 0.2.3 decoder runs as a chain of ordinary GPU steps, so it behaves the same on any Apple Silicon GPU and under any load from other apps. Its output is token-identical to 0.2.2.
• If the Metal decoder ever fails, Phonon switches to MLX for the rest of the session instead of retrying on every dictation.
• Everything else is 0.2.2. Same models and weights; updating downloads nothing new.

Known issues

• Connecting or disconnecting Bluetooth headphones (for example AirPods) while Phonon is idle can crash it. Reopen Phonon; a fix is being tested.
• Tested on M4 and M4 Max. M1 to M3 have not been tested on hardware yet; if the Metal engine fails its startup self-test there, Phonon uses MLX as before.
• English only.

Install

brew install --cask infatoshi/phonon/phonon

or download Phonon.dmg below (signed and notarized).

License

Phonon stays GPL-3.0. The refiner weights are under the LFM Open License v1.0. Parakeet is CC-BY-4.0 (NVIDIA). The engine links llama.cpp (MIT, commit 39febd2 plus the patch in third_party/llama.cpp).

Phonon 0.2.2

Phonon 0.2.2 (macOS)

Hands-free dictation now works on the Option key too.

What changed

• Double-tap to keep recording, on every hold key: hold the key to talk and release to insert, or double-tap it to keep recording without holding it, then tap once to stop. This already worked on the Globe (fn) key; it now works on the Option shortcut as well, for keyboards without a Globe key. Settings shows the hint under the shortcut picker.
• Everything else is 0.2.1: streaming correction while you talk, speech recognition on Phonon's Metal kernels, and the faster drafter. Same models and weights; updating downloads nothing new.

Known issues

• Connecting or disconnecting Bluetooth headphones (for example AirPods) while Phonon is idle can crash it. Reopen Phonon; a fix is being tested.
• The Metal speech kernels are untested on base M1 and M2 chips. If they fail the startup self-test there, Phonon uses MLX as before.
• English only.

Install

brew install --cask infatoshi/phonon/phonon

or download Phonon.dmg below (signed and notarized).

License

Phonon stays GPL-3.0. The refiner weights are under the LFM Open License v1.0. Parakeet is CC-BY-4.0 (NVIDIA). The engine links llama.cpp (MIT, commit 39febd2 plus the patch in third_party/llama.cpp).

Phonon 0.2.1

Phonon 0.2.1 (macOS)

Phonon now transcribes and corrects while you talk, so most of the work is done before you let go of the key.

What changed

• Streaming: while you hold the key, Parakeet re-transcribes the recording every 0.4 s and the refiner corrects the transcript so far. At key release only the last few seconds are transcribed again, and the refiner checks its draft instead of starting over. On an M4 Max, the median time from key release to corrected text went from about 215 ms to about 30 ms (40 real dictations, measured inside the app's engine; pasting adds a few ms). If you pause before letting go, the last pass already heard every word: those dictations were ready in about 12 ms.
• Same text as before: streaming gives the same transcript as transcribing the whole recording at once, within measurement noise (1,351 public clips with human transcripts: +0.0005 word error rate, 95% interval -0.0010 to +0.0019).
• Streaming is on by default on Macs with 16 GB of memory or more and off below, where it would keep the GPU busy while you talk. "stream_refine": true or false in ~/Library/Application Support/Phonon/settings.json overrides it.
• Speech recognition runs on Phonon's own Metal kernels instead of MLX: the same Parakeet weights, read directly from the existing download, with no new files. If the kernels can't run on a Mac, a startup self-test switches to MLX.
• A faster drafter for the refiner (G16): when streaming is off, the correction step is about 1.2x faster (185 to 148 ms median on an M4 Max, 40 dictations), with identical output. It replaces the drafter file in the model download (about 215 MB).

Known issues

• Connecting or disconnecting Bluetooth headphones (for example AirPods) while Phonon is idle can crash it. This also happens in 0.2.0. Reopen Phonon; a fix is in progress.
• The Metal speech kernels are untested on base M1 and M2 chips. If they fail the startup self-test there, Phonon uses MLX as before.
• English only.

Install

brew install --cask infatoshi/phonon/phonon

or download Phonon.dmg below (signed and notarized). Updating from 0.2.0 downloads only the new drafter file; the other weights are unchanged.

License

Phonon stays GPL-3.0. The refiner weights are under the LFM Open License v1.0. Parakeet is CC-BY-4.0 (NVIDIA). The engine links llama.cpp (MIT, commit 39febd2 plus the patch in third_party/llama.cpp).

Phonon 0.2.0

Phonon 0.2.0 (macOS)

Phonon's correction model is now one I trained myself, on public audio only. It replaces Gemma 4 E2B as the default.

What changed

• New default correction model: Phonon refiner 1.7B, fine-tuned from LiquidAI's LFM3-1.7B. The training audio is public YouTube technical talks, transcribed by Parakeet and labeled by DeepSeek V4.1 Flash running locally. No one's personal dictations went into it. Weights: https://huggingface.co/Infatoshi/phonon-refiner-1.7b
• Accuracy: 28-31% fewer word errors than Gemma 4 E2B on two held-out dictation sets (12.1% vs 16.9% and 10.8% vs 15.7%, measured on the shipped 4-bit model).
• Speed: the refiner runs on Phonon's own Metal engine with a speculative drafter, bundled in the app. Key-up to text is about 0.3 s (median, dictations under 30 s) on an M4 Max and about 1.3 s on an 8 GB M1. If the engine can't start on a Mac, Phonon loads the same refiner on MLX instead and says so in the startup window.
• Smaller first download: about 4.9 GB of weights (Parakeet 2.47 GB + refiner 2.40 GB), down from 5.6 GB, plus about 0.17 GB of Python packages. The startup window shows real download progress.
• The engine is built for macOS 14 and later, on Apple Silicon.
• Your dictionary still feeds the refiner on every request (the "Vocabulary the speaker may use" line).
• Also since 0.1.7:
  • Learn from your edits: when you fix a word Phonon typed, it offers to add the word to your dictionary. Nothing is added until you approve it.
  • Private apps and sites: Phonon reads nothing in them, so no screen reading and no edit learning. Signal and Discord are on the list by default; password prompts and password managers are always private.
  • Screen context reads only the windows you can see, plus any you switch to while speaking, with on-device OCR. Screenshots are never saved.
  • Meeting notes (off by default): when a call starts in Zoom, Teams, FaceTime, Slack, Discord or a browser meeting, Phonon asks whether to record your mic and the call audio to local WAV files.
  • History shows what Phonon wrote next to what you kept.
  • Fixed a crash when a call app switched audio devices mid-session.
  • The first dictation after launch no longer waits on screen reading, and the engine warms up when you press the key.
  • Settings use plain wording, with switches that read on or off at a glance.

License

Phonon stays GPL-3.0. The refiner weights are under the LFM Open License v1.0, which allows commercial use only for organizations under $10M in annual revenue. Parakeet is CC-BY-4.0 (NVIDIA). The engine links llama.cpp (MIT, commit 39febd2 plus the patch in third_party/llama.cpp).

Install

brew install --cask infatoshi/phonon/phonon

or download Phonon.dmg below (signed and notarized). The Homebrew formula builds from source and corrects on the MLX refiner; the cask and the DMG include the native engine.

Known limits

• The public refiner knows new technical terms less well than a refiner trained on your own dictations: on my held-out sets it got 49% of new terms right, against 89% for my personal model. Adding terms to the dictionary is the lever for now.
• English only.
• Windows beta is unchanged and still uses Gemma 4 E2B.

Phonon 0.2.0 Windows beta 1

Beta. No one has run this on real Windows hardware yet. Phonon is looking for one tester.

Continuous integration builds this executable on windows-latest and runs one
real dictation through it on every push: assets/startup.wav goes through the
real Parakeet model and the real Gemma correction model, and the run fails
unless the transcript and the correction both pass the same rules the macOS
readiness gate applies. That gate is green for this build.

The run that produced this file

raw:      "Hello Fluid Voice."
polished: "Hello Fluid Voice."
asr 2583 ms, correction 16266 ms

That correction time is a two-core shared runner. A desktop CPU should be several
times faster. Nobody has measured one.

Install

1. Download phonon-win.exe. It is one file. There is no installer.
2. Windows SmartScreen shows Windows protected your PC, because the build is
   not code signed. Choose More info, then Run anyway.
3. First run downloads 3.8 GB into %LOCALAPPDATA%\Phonon: the sherpa-onnx
   tools, the llama.cpp CPU build, Parakeet TDT 0.6b v2 as int8 ONNX, and
   Google's quantisation-aware Gemma 4 E2B q4_0 GGUF. Every file is pinned to a
   release tag or a commit and is checked against a SHA-256 built into the
   executable. The tray tooltip carries the progress.
4. Everything after that is local. No account, no cloud, no network.

Use

Hold Right Ctrl and talk. Let go and the text lands in whatever window has
focus. Double-tap Right Ctrl to latch a long dictation, then press it again to
stop. Right-click the tray icon for the menu.

Phonon swallows Right Ctrl while it runs, so Right Ctrl stops acting as a
modifier. Left Ctrl is untouched. PHONON_WIN_HOTKEY takes rightctrl,
leftctrl, rightalt, rightshift, capslock, or f13.

phonon-win.exe info      # what first run downloads, and what is installed
phonon-win.exe fetch     # download and verify, then stop
phonon-win.exe selftest  # one real pass through both models

Requirements

Windows 10 or 11, x64. Everything runs on the CPU. 16 GB of memory is
comfortable and 8 GB is the floor. About 4 GB of disk for the models.
Unpacking uses the tar that ships with Windows 10 build 17063 and later.

What is missing against macOS

Windows runs the recognise-and-correct path only. There is no microphone
priority list, no streaming preview, no screen context, no dictionary
retrieval, and no corpus. Recognition loads the encoder for every pass, so the
first word costs more than it does on the Mac.

If it breaks

Run phonon-win.exe selftest from a terminal and open an issue with what it
prints. Deleting %LOCALAPPDATA%\Phonon starts over.