Voxden

Dictation for Windows and Mac that runs on your own computer.
Press a key, talk, and the text lands in whatever app you were typing in.
Free for 3,000 words a week. Your audio stays on your computer unless you turn on Voxden Cloud.

⬇ Download for Windows or Mac

Why Voxden

Tools like Wispr Flow and Dragon do this well, but they charge a monthly fee and send your voice to a server. Voxden does the same job on your own machine, on Windows or a Mac. The Free plan covers 3,000 words a week; Voxden Pro removes the limit and adds cloud dictation.

• Works in every app. Chat, email, the browser, your IDE, a terminal. If you can type there, you can dictate there.
• Private by design. Speech recognition, cleanup and grammar fixes run locally. Your audio leaves the PC only if you turn on Voxden Cloud, a Pro feature that is off by default.
• Real speech models, not the Windows built-in one. Pick from Parakeet, Whisper large-v3 turbo, Whisper large-v3 or Qwen3-ASR. Strong on accents. On your PC they dictate in English; Voxden Cloud covers 18 languages, Hindi and Hinglish included, up to three at once.
• Learns your words. Correct a transcript once and Voxden remembers the name, the product, the jargon. Whisper and Qwen3-ASR are given those terms before they decode.
• Knows where you are typing. Casual in WhatsApp, formal in Outlook. Fillers like "um" and "you know" are removed, spoken numbers become digits.
• Nothing to set up. One installer and a one-time sign-in. No Python, no API keys, no Hugging Face account.

How it works

_{Island, the default flow bar: at rest, open on hover, and while recording.}

_{Orb, the other style in Settings → Display: at rest, open on hover, and while recording.}

1. Click where you want the text. Any text box in any app.
2. Press Ctrl + Win (on a Mac, Cmd + Shift + Space) and talk. The flow bar at the bottom of the screen opens and shows your voice level.
3. Press the same keys again. Voxden transcribes, cleans up the sentence and pastes it where your cursor was.

That is the whole workflow. Esc cancels without pasting. Ctrl + Alt + V pastes your last dictation again. Open Settings → General → Shortcuts to change either shortcut; an install updated from an older version keeps the shortcut it had. General also contains your microphone, dictation languages, app language, dictation mode, and dictation speed.

Between dictations the flow bar rests at the bottom of the screen as a small black pill. Hover it and it opens into settings, the microphone and screenshot capture. Click it to dictate, or drag the bar itself anywhere on any monitor. It remembers where you left it.

Install

Windows

1. Download Voxden-Setup-<version>.exe from the latest release and run it. Windows 10 or 11, 64-bit.
2. If Windows shows a blue "Windows protected your PC" screen, click More info, then Run anyway. Voxden is not yet code-signed, so Microsoft SmartScreen flags the installer as unfamiliar. It is not a virus report. If your browser also blocks the download, choose Keep from its download menu.
3. On first launch, sign in with Google or an emailed code, then choose a speech model. Voxden downloads only that one and checks the file hash. You can switch models later in Settings → Speech engines.
4. Start dictating.

Mac

1. Download Voxden-<version>-mac-arm64.dmg from the latest release, open it and drag Voxden into Applications. macOS 12 or later on Apple silicon (M1 or later). There is no Intel build.
2. The Mac app is not yet signed with an Apple certificate, so macOS says it cannot verify Voxden. Click Done, open System Settings → Privacy & Security, scroll to Security and click Open Anyway. On macOS 14 and earlier, right-click Voxden and choose Open instead. Step by step, with pictures, on voxden.app/download.
3. Sign in and choose a speech model as on Windows, and allow the microphone and Accessibility when macOS asks. On a Mac the local models are Parakeet v3 and Whisper; Qwen3-ASR and the GPU packs are Windows-only. Voxden Cloud on Pro works the same on both.
4. A Mac does not update itself yet. Voxden tells you when a newer version is out; download the new disk image and drag it over the old copy.

Model	Download	Best for
Parakeet v3 (default)	670 MB	Fastest, and small enough that a fresh install dictates within minutes.
Whisper large-v3 turbo	1.6 GB	Whisper at half the download. Uses the words in your dictionary and is lighter than large-v3 on a CPU.
Qwen3-ASR 1.7B	4.7 GB	Best at catching names and the words in your dictionary.
Whisper large-v3	3.1 GB	Uses the words in your dictionary, smaller download than Qwen. The only model that can use a model trained on your voice.

Dictation on your PC is in English with every model. Hindi, Hinglish and other languages need Voxden Cloud on a Pro plan.

Any modern PC runs Voxden on the CPU. An NVIDIA card makes Qwen and Whisper several times faster with an optional download from Settings. See Speed it up with your graphics card.

_{Help → Setup guide inside the app walks a new user through the same steps.}

Free and Pro

Voxden asks you to sign in once, before your first dictation, with Google or a six-digit code sent to your email. There is no password to remember.

• Free dictates up to 3,000 words in any seven days, on the speech model you downloaded. The seven days start with your first dictation, and the words come back when they end. Dictation works offline once your model is downloaded.
• Voxden Pro is a monthly subscription. It removes the weekly word limit and adds Voxden Cloud. Upgrade, manage or cancel it in Settings → Plans & billing. Prices for your region are on voxden.app/pricing.
• Voxden Cloud stays off until you choose it in Settings → Speech engines → How Voxden listens. Your audio is then sent to Voxden's service to be transcribed: no model download, and 18 languages, Hindi and Hinglish included, up to three at once. Pro includes 900 cloud credits a month and 1,200 in your first month; one credit is one minute of audio. If the cloud is too busy to answer, the model on your PC transcribes instead, if you have one installed.

Features

Capture: point, speak, and attach

Hover over the flow bar and click the capture-corners icon beside the microphone, or choose Capture screenshot… from the tray menu. Drag a region on any monitor, then use Circle, Arrow, Pen, or Hide to mark the detail directly on the screenshot. Listening starts automatically, so you can speak while marking. Press your normal dictation shortcut to stop; Voxden pastes your words and the marked screenshot into your chat box, then dismisses the capture. The small toolbar also has a stop button.

Capture remembers the app you were using. To use another destination, click its message box before stopping dictation. The destination must support pasting images. Voxden never sends the message automatically in Capture mode. If a paste fails, the capture stays available and your shortcut retries it.

Press Escape to cancel, or use the retake icon to start a fresh capture. Screenshots and annotations stay in memory until pasted or cancelled. Speech is transcribed on your PC unless Voxden Cloud is on. Each selection stays within one monitor; captures do not span displays.

Dictionary: teach it your words once

Edit any transcript in the history and Voxden learns the correction. The next time you say "cooper netties" it writes Kubernetes. Whisper and Qwen3-ASR are handed the terms before they decode, so they come out right the first time rather than being patched afterwards; Parakeet has no way to take them, so on Parakeet the dictionary corrects the transcript afterwards. You can also add words by hand.

With Settings → General → More options → Auto-add to dictionary enabled (the default), correct a recently dictated word directly in a supported Windows text field. After you pause typing, Voxden adds the corrected spelling and shows Added “Kubernetes” to dictionary in the flow bar, with Undo. Undo removes that addition without changing your text. Automatically learned spellings guide the speech model; replacement rules can still be added explicitly in Dictionary.

This runs locally and watches only the field that received the latest dictation, for up to 90 seconds. Switching fields/apps, clearing the field, starting another dictation, or turning the setting off ends observation. Password, read-only, unsupported and very large fields are skipped, as are dictations sent automatically. Field contents stay in memory; only learned terms are saved. History edits still offer replacement suggestions for review in Dictionary.

Writing style: casual in chat, formal in email

Voxden looks at which app is in front and picks a tone for it. Personal messages, work messages and emails each get their own setting. A tone sets capitals and punctuation and keeps your words: Formal adds a closing full stop, Very casual writes in lower case without one. Filler words are removed, repeated words collapsed, spoken punctuation ("comma", "new paragraph") applied, and numbers written as digits. Turn on Verbatim mode to paste exactly what you said.

Enable Auto cleanup in Writing style for lightweight English proofreading with no extra model or download. It fixes common agreement and verb mistakes ("we was gonna go" → "We were gonna go."), punctuation spacing, sentence casing, and missing end punctuation. It preserves casual wording and applies your selected tone afterwards, so Very casual still uses lowercase and omits a final period. The option starts off, pauses in Verbatim mode, and leaves other dictation languages unchanged. It uses local rules; it does not restructure long sentences or resolve ambiguous grammar.

Insights: see how much you actually talk

Words per minute against typing speed, time saved, which apps you dictate into, a streak calendar, and milestones on a bookshelf. All computed locally from your own history.

Settings

Choose whether Voxden listens on your PC or through Voxden Cloud, pick the speech model and the processor it runs on, change the shortcut, switch between toggle and push-to-talk, and decide whether other audio is silenced while you dictate.

In Settings → Data and privacy, use Delete beside Keep recordings to clear saved dictation audio. Transcripts, training clips, exported WAV files, and your choice to keep future recordings are preserved.

And the rest

• Mute other audio while dictating. Spotify and other Windows media sessions pause, while calls, videos, games, and other playback are silenced until the microphone closes. Music or output you muted yourself stays that way.
• Spoken numbers become digits. "twenty five percent" → 25%, "version one point zero point sixteen" → version 1.0.16, "the twenty fifth" → the 25th. Small bare numbers stay words where style guides want them ("two cats").
• Voice commands. new line, new paragraph, period, comma, question mark, scratch that.
• A failed dictation keeps its recording. If a dictation never becomes text, for example because Voxden heard only silence, no speech model was set up yet, or Voxden closed mid-dictation, the recording waits on the Dictation page under Not transcribed for 14 days. Recover turns it into text with the speech engine you use now. This follows Keep recordings in Settings → Data and privacy.
• What's new bell. New engines and features are announced inside the app. Bug fixes are deliberately not announced.
• Train on your own voice. Optionally keep the audio behind dictations you correct, then fine-tune Whisper on it. Off by default, nothing is uploaded. See Training on your own voice.

Voxden vs Wispr Flow

	Voxden	Wispr Flow
Price	Free for 3,000 words a week, or Voxden Pro; MIT licensed	Subscription
Where speech is processed	Your PC, or Voxden Cloud if you turn it on (Pro)	Their servers
Account required	Yes: Google or an emailed code	Yes
Works offline	Yes, after sign-in and the model download	No
Platform	Windows, macOS (Apple silicon)	Windows, macOS, iOS
Speech models	Parakeet, Whisper, Qwen3-ASR, your own fine-tune, or Voxden Cloud	Cloud
Custom vocabulary	Yes, fed to Whisper and Qwen3-ASR before decoding	Yes
Per-app tone	Yes	Yes

If you need your phone, or a Mac with an Intel chip, Wispr Flow is the better fit today. If you want the same thing on Windows or an Apple silicon Mac with your voice kept on your own computer, that is what Voxden is for.

Speed it up with your graphics card

You can skip this. Voxden works on any PC without it. If dictation feels slow, one extra download can speed up your model. Find these in Settings → Speech engines.

Your hardware	What to do
NVIDIA card + Whisper	Choose Download 553 MB under Speed up with your GPU
NVIDIA card + Qwen3-ASR	Choose Download under Speed up with your GPU
AMD card + Qwen3-ASR	Choose Download under Speed up with your GPU. Only some AMD cards support it.
AMD or Intel graphics + Parakeet	Set Processor to AMD or Intel GPU under Settings → Speech engines → Advanced. This needs a separate 2.5 GB download.
Anything else	Leave Processor (Settings → Speech engines → Advanced) on Auto.

Each pack speeds up one kind of model; the Whisper pack covers both Whisper models. On a strong CPU the gain can be small: on a 24-thread part Parakeet measured 17x realtime on the CPU against 15.9x on DirectML, so the GPU matters most where the CPU is the weak part.

Run from source

npm install
npm start

Uses the system Node install. A source checkout picks up a Python environment in this order: VOXDEN_PYTHON, the downloaded speech engine if you installed one, .venv/Scripts/python.exe, then the system Python. Tests run with npm test.

Useful environment variables for development:

Variable	Effect
VOXDEN_PYTHON	Point at a Python environment you maintain yourself, including a CUDA PyTorch one. Packaged builds never search system Python.
VOXDEN_DEVICE	cpu, cuda, directml or auto. Overrides the UI setting; unsupported backends fall back to CPU.
VOXDEN_CPU_THREADS	Override the CPU thread count (default: half the logical processors, capped at 16).
VOXDEN_MODEL	Use a specific model directory, for example your own fine-tune.
VOXDEN_LAZY_ASR=1	Defer loading the model until the first dictation instead of at startup.

Technical notes

Everything below is here for the curious and for contributors. None of it is needed to use the app.

Speech engines and the bundled runtime

The Windows installer includes a self-contained speech runtime with Whisper, Qwen3-ASR, Parakeet, CPU PyTorch, and DirectML. End users do not install Python, run pip, or need a Hugging Face account.

On first launch, the welcome dialog downloads the model you choose and nothing else. Parakeet v3 (~0.7 GB) is preselected. Download and use on any other row in Settings → Speech engines downloads that model and switches to it in one click: Whisper large-v3 turbo (~1.6 GB), Qwen3-ASR 1.7B (~4.7 GB) or Whisper large-v3 (~3.1 GB). The float32 Parakeet weights (~2.5 GB) are a separate optional download for AMD and Intel graphics. Existing app model caches are verified and reused where possible. Setup checks SHA-256, resumes interrupted downloads, and keeps completed models across updates.

Starting the app, switching engines, and dictation never download models in the managed runtime. Removing speech engines stops their processes and disables dictation; the window, history, and settings still work. Download again to reinstall. A normal launch opens the dashboard; launching with Windows stays in the tray.

Settings → Speech engines lists four local models, each with one button. Switching restarts the sidecar and releases the previous model before loading the next one.

Voxden Cloud (Settings → Speech engines → How Voxden listens, off by default, Pro only) sends each dictation's audio to Voxden's account service, which forwards it to a hosted speech model and meters the seconds against the plan's monthly cloud credits (one credit is one minute of audio). The app waits about 25 seconds for an answer, and up to 40 for a long clip. If the speech provider is too busy to answer, the dictation is transcribed on this PC instead when a local model is installed, and the flow bar says "Cloud busy. Using this PC…" while it switches. Any other failure (too slow, unreachable, over the cap, or an account that is not Pro) fails the dictation with the reason shown on the Voxden Cloud card. Switch How Voxden listens to On this PC to dictate locally. Audio leaves the PC only while this is on. The service and its relay live in server/.

• Parakeet v3 (TDT 0.6B) — the default on a fresh install; small and fast. When Whisper or Qwen is selected, Dictation speed Fast (and Auto in chat apps such as ChatGPT, Claude, Slack, Discord, WhatsApp) still uses Parakeet for lower latency. If Parakeet is missing, Fast uses the selected engine with a cheaper decode.
• Whisper large-v3 turbo — Whisper at half the download, on the same faster-whisper runtime and with the same dictionary support as large-v3. It downloads from Hugging Face through the same catalog as Qwen and Parakeet.
• Qwen3-ASR 1.7B — the download for names and dictionary terms, which it caught most often of the three models measured (measurements), through the official qwen-asr Transformers backend. A settings file from before the engine picker existed keeps whichever of Qwen or Whisper is already downloaded rather than reverting to Parakeet.
• Whisper large-v3 — installed through faster-whisper; the mature alternative with word timings and confidence scores. CUDA float16 where available and CPU int8 otherwise.

This build includes CPU PyTorch, so Qwen works without extra downloads. Optional Qwen CUDA acceleration (NVIDIA) and Qwen ROCm acceleration (only AMD GPUs on AMD's Windows PyTorch list) are separate downloads. The Whisper cuBLAS pack does not accelerate Qwen. DirectML accelerates Parakeet only. Settings says Qwen3-ASR is using your GPU only after the sidecar verifies it, whatever Processor is set to.

The CPU path runs on half the logical processors, capped at 16. CTranslate2 uses four on its own whatever the machine has, which on a 12-core part measured 1.5x realtime against 4.2x with the cores it actually had.

Maintainer instructions for the speech engine and model release assets are in docs/ASR_ASSETS_RELEASE.md.

AMD and Intel graphics (DirectML)

Settings → Speech engines → Advanced → Processor offers AMD or Intel GPU, which runs Parakeet on ONNX Runtime's DirectML provider. DirectML targets DirectX 12 rather than a vendor, so one option covers Radeon, Intel integrated graphics and Arc.

Auto does not pick it. It stays CUDA-or-CPU. Nearly every PC has a DirectX 12 card, so ranking DirectML above the CPU would move most users onto a 2.5 GB download in place of a 0.7 GB one for a gain they may not have: on a 24-thread CPU the two measured 15.9x against 17.0x realtime. DirectML earns its place where the CPU is the weak part, which is a thing the person at the machine knows and auto does not.

Only Parakeet has that DirectML path. CTranslate2 has exactly one GPU backend and it is CUDA. Qwen3-ASR on AMD uses CPU PyTorch unless the GPU is on AMD's Windows ROCm PyTorch list and the separate Qwen ROCm pack is installed and verified. That list is short: it is not every Radeon. Whisper on a Radeon stays on the CPU. Settings offers no GPU speed-up for Whisper on a Radeon.

The GPU path drops quantization: DirectML gets the float32 weights (2.5 GB) rather than the int8 ones (0.7 GB), because the int8 build is a QDQ graph whose quantize/dequantize pairs cost a GPU more than they save. Measured on one DirectX 12 card, int8 on DirectML ran at 6.9x realtime against 15.9x for float32 and 17x for int8 on a 24-thread CPU. Each precision keeps its own directory under the model folder, so moving the setting between the CPU and a GPU does not throw the other download away.

DirectML arrived in speech engine asr-win-x64-v2. An older install has no DirectML provider, and Voxden says so and asks for a reinstall from Settings. On your own Python, pip install onnxruntime-directml instead of onnxruntime or onnxruntime-gpu. One ONNX Runtime build per environment, never two.

Startup, model loading and the paste helper

On launch, Voxden probes the speech runtime (a cheap import check) and then loads the selected model in the background a moment later, at below-normal process priority and with its CPU threads capped to half the logical processors, so the first dictation does not wait for a multi-gigabyte load and the desktop stays responsive while it happens. The engine also runs a short silent clip through itself before reporting ready, so the first real dictation is answered at full speed. Set VOXDEN_LAZY_ASR=1 to defer the load until the first dictation instead.

The Windows helper that reads the foreground window, pastes, pauses music, and silences other playback runs as a small pool of long-lived PowerShell processes. Earlier builds started a new process for every call, and each one compiled the helper before answering, which cost about a quarter of a CPU second twice a second for the life of the app and put the paste a full second behind the transcript.

Mute other audio while dictating pauses supported Windows media sessions and mutes active playback devices before recording starts. This silences ordinary audio such as Discord calls even when it has no media controls. When the microphone closes, Voxden unmutes only devices it muted and resumes only media it paused; pre-existing mute and pause states stay untouched. Rapid dictations, cancellation, errors, and quit share the same ordered restore flow.

The flow bar

Island is the default flow bar: one black capsule with a hairline rim, and no glow or shadow in any state. With Always show flow bar on (Settings → Display → Behavior & motion, on by default) it rests at the bottom of the screen as a small pill. Hover it and it opens into settings on the left, the microphone in the middle and screenshot capture on the right. Click the bar to dictate, the gear to open Voxden's settings, or the screenshot button to start a capture. There is no drag handle: press the bar itself and drag to move it anywhere on any monitor. Where you drop it is remembered.

Orb, a glowing voice sphere with a grip beside it for dragging, is the only other style, under Settings → Display → Flow bar style. A settings file that still names the retired Classic or Ribbon style gets Island.

The bar keeps its position across restarts. If the monitor it was on goes away, it moves to the nearest one that is still there and returns when that screen comes back. Once you have moved it, Reset position in Settings → Display → Behavior & motion puts it back at the bottom of your main display. Wherever the bar sits, dictation still pastes into the window that had focus, not into the screen the bar happens to be on.

Dictionary and vocabulary

Edits in history teach a local dictionary (data/dictionary.json). Future transcripts apply those replacements before paste (case-insensitive, longest phrase first, in any script). Learned phrases are listed on the Dictionary page; delete one with its bin icon.

Dictionary terms are also given to the speech engine before it decodes, through whatever input that engine actually has: Whisper takes them as initial_prompt, Qwen3-ASR as its context system message. Parakeet has no such input at all, so on that engine the dictionary is applied to the transcript afterwards and Voxden says so rather than pretending otherwise. Terms are ranked by how recently they were added and used and packed into each engine's own token budget, so a word you add now is in the very next dictation even if your dictionary is far larger than any prompt window.

An "accurate" dictation keeps its dictionary. Voxden will still switch to the faster engine on a CPU when there is nothing to lose, but not when that would mean dropping the terms you taught it.

Full detail, including how any of this is measured, is in docs/VOCABULARY_AND_ACCURACY.md.

Cleanup and numbers

A tone changes capitals and punctuation, never words. Every tone removes the sounds "um", "uh" and "hmm", and "you know" only where punctuation marks it off as an aside ("I was, you know, thinking"). "Like", "I mean", "kind of" and "sort of" stay, because they can carry meaning, and sentences such as "Do you know the answer?" and "I like this design" are never touched. Repeated words collapse only when they are a stumble ("the the", "how can I, how can I start"); "very, very good", "bye bye" and a sentence said twice stay as spoken.

Spoken numbers are written as figures: "one point zero point sixteen" becomes 1.0.16, "twenty five percent" becomes 25%, "twenty twenty six" becomes 2026, "the twenty fifth" becomes the 25th, and "five five five one two three four" becomes 5551234. A bare "one" to "nine" stays a word ("one of them", "two cats") unless a unit or a label makes it a figure ("five percent", "page three", "version two"), which is what style guides ask for. Write numbers as digits on the Writing style page turns this off. Verbatim mode never rewrites numbers.

Notifications

The bell at the right of the title bar carries what is new: a new engine, a new language model, a feature that did not exist before, and an update that has finished downloading. A count under the bell says how many you have not looked at; opening the panel clears it. Rows stay until you dismiss one with x or use Clear all.

Bug fixes are deliberately not announced. A fresh install is told only what shipped in the version it installed, and updating tells you what landed in every version you skipped.

Announcements ship inside the app, in the catalog at the top of src/announcements.js. Add an entry with since set to the version it ships in; the id is permanent, because a dismissed id is remembered so it cannot come back.

Training data and training on your own voice

Off by default. Turn on Settings → Data and privacy → Keep audio for training and Voxden keeps the recording behind any dictation you correct, paired with your corrected text. That pair is the only ground truth this app ever gets, and it is normally deleted the moment transcription returns.

Clips you never correct age out of a small pending window. Corrected ones land in data/audio/corpus/ with a data/audio/pairs.jsonl manifest. Deleting a dictation deletes its recording; turning the setting off deletes all of them. Nothing is uploaded.

To see what has accumulated:

node scripts/export-training-data.js

Add --write to emit train.jsonl and eval.jsonl with absolute paths and a sentence key, ready for a Whisper fine-tune.

Once enough clips have accumulated, training/ holds a LoRA fine-tune of Whisper on them, a CTranslate2 conversion step, and an evaluation that compares the result against stock large-v3 on held-out clips. See training/README.md.

A finished model lands in models/voxden-tuned/ and the app picks it up on its own; Settings → Speech engines → Advanced gets a Use your tuned model toggle while Whisper large-v3 is in use. VOXDEN_MODEL overrides both.

Contributing and feedback

Found a bug, or a phrase Voxden keeps getting wrong? Send it from Help → Feedback or bug report in the app, or open an issue. Pull requests are welcome. If Voxden saves you time, a star on the repo helps other people find it.

License

MIT. Third-party notices are in THIRD_PARTY_NOTICES.md.