English · Русский · macOS version →

Voica for Windows

A Windows tray app for voice dictation with punctuation, powered by Groq Whisper.

---

Press a hotkey, speak, and Voica inserts clean, punctuated text into whatever field you're typing in. Bring your own Groq API key.

Voica is a tiny background app that lives in the system tray. It's a native Windows (C# / .NET 8 / WPF) implementation of Voica (macOS), and follows the same cross‑platform behavior spec.

Support the project

Voica is free and stays free — every feature, no subscription. If it saves you time, you can back the work on Boosty: the road to 1.0 for both the Windows and the macOS version.

Features

• Global hotkey dictation — Push‑to‑talk (hold) or Toggle (press to start/stop). Default: Toggle + Right Alt. Pick a preset key (Right/Left Alt, CapsLock, ScrollLock, Pause) or record a custom combination (e.g. Ctrl+Shift+Space). In Toggle mode a double tap starts the recording, so a stray press can't.

• Dictation bar — a floating capsule at the bottom of the screen shows the live level while you speak, with × to cancel (the audio is discarded) and ✓ to stop and transcribe. It never takes focus, so the text still lands in the field you were typing in. Can be turned off, and then the tray icon indicates the state instead.

• Punctuation via Groq Whisper — pick the model (whisper-large-v3-turbo or whisper-large-v3) and the language (auto‑detect, great for mixed Russian/English, or force Russian/English).

• Local offline engine (optional) — recognition fully on your PC via GigaAM v3 (Russian, with punctuation), no network and no API key. If the cloud is selected but the network is down, Voica automatically falls back to the local engine when the model is installed. Trade-off: Latin words can come out as a mix of alphabets (Dпсик instead of DeepSeek) — that is exactly what the vocabulary fixes, see below. The recognition hint stays cloud-only.

• Auto‑insert into the focused field (synthesized Ctrl+V), and the text is always copied to the clipboard as a fallback. Or show an editable result window.

• History (SQLite) — browse, search (Ctrl+F — it also looks at what the engine heard before the corrections, so you find a dictation by what you said), re‑copy, play the audio, delete (one or a whole multi‑selection), and export the entire history to Markdown, CSV or JSON.

• Audio retention — keep recordings for N days (default 30; 0 = keep forever), or don't store audio at all.

• Vocabulary — list the terms recognition mangles (names, jargon, anglicisms). It works in three layers:

  • Rules, right on your PC. Garbled spellings are pulled back to the ones you listed — no key, no internet, both engines — and it switches itself on whenever the vocabulary isn't empty. This is what makes the local engine genuinely self-contained.
  • A recognition hint — cloud only: the list goes to Whisper and biases what it hears. Soft limit ~800 characters (the model only reads the last ~224 tokens; a live counter in Settings shows the budget), so keep the terms that matter at the end of the list.
  • An AI pass (Groq LLM), optional — handles what rules cannot: grammatical case and badly garbled terms. Needs the key and internet; if the request fails you keep the text the rules produced, so it never makes things worse.

  More: Словарь терминов и ИИ-исправление (in Russian).

• Text clean-up — two more rules that run without a key or a network, both on by default and both switchable:

  • Filler sounds go — the drawn-out "э-э-э", "ммм", "хмм" that mean nothing in speech and clutter the text. The rule reads the shape of a word, not a list of spellings, so it catches every way recognition happens to spell a mumble. A real word that was merely drawn out is straightened rather than dropped ("ну-у-у" → "ну"), and numbers, abbreviations and millimetres are left alone. Turn it off if you transcribe speech verbatim — it deletes what was said.
  • Quotation marks are repaired — straight quotes become guillemets by position, a missing space after a colon comes back, and unpaired quotes are removed. Recognition places quotes however it happens to, and the AI pass adds its own; English text is left alone.

• Works behind a corporate proxy — Voica uses the proxy Windows is configured with and authenticates as the signed‑in user (domain SSO), so no password is typed into the app or stored by it. The switch is on the Network tab, next to a line saying which route requests actually take; turning it off goes straight out. If a proxy refuses, the message names its address.

• Update checks against this repo's GitHub releases (opt‑in, once a day). Voica never downloads or installs anything itself — it just opens the release page.

• Privacy — no backend, no telemetry. Network is used only for Groq (cloud transcription / AI correction) and GitHub (optional update checks, one‑time model download). With the local engine, audio never leaves your PC. Your API key is stored encrypted with Windows DPAPI.

• English & Russian UI, by system language.

Screenshots

Settings — six tabs: engine and API key, dictation, vocabulary, data, network, about.

Dictation bar — while recording (level wave, cancel, stop):

History — the list on the left, the record's text on the right; search covers what was said before any fixing:

Tray icon (idle):

Requirements

• Windows 10 version 1809 (build 17763) or later, x64.
• A Groq API key for cloud recognition — create one at https://console.groq.com (free tier's whisper-large-v3-turbo). Not needed if you use the local offline engine.

Install

Take the installer from the latest release:

• Voica-Setup-<version>.exe — the normal way in. It carries the self‑contained build, so there is nothing to choose and nothing else to install. It sets Voica up for your user only (%LocalAppData%\Programs\Voica), which means no UAC prompt, adds a Start‑menu entry, and registers a proper uninstall record. Installing over a running copy offers to close it first. Uninstalling removes the program but keeps your data — history, settings and the API key stay in %APPDATA%\Voica (wiping those is the app's own Delete all data, which asks you to type a confirmation).

The two bare executables stay on the release page for anyone who prefers to run one file with no installation at all — you just have to know which:

• Voica.exe (~80 MB) — fully self‑contained, nothing else to install.
• Voica-fx.exe (~37 MB) — smaller, but needs the .NET 8 Desktop Runtime installed once.

Either way Voica runs in the system tray (no main window), and either way your data lives in %APPDATA%\Voica, so you can switch between them without losing anything. See below about the SmartScreen warning — it applies to the installer too.

Why does Windows warn about this app?

Voica isn't code‑signed yet, so on first run SmartScreen shows "Windows protected your PC." Click More info → Run anyway. This is expected for an independent, unsigned app — not a sign that something is wrong:

• Voica is open source, and every release is built by GitHub Actions from the tagged commit (release author github-actions[bot]), not on anyone's personal machine — so the binary is reproducible from the source you can read.
• You can verify your download: the release page shows a sha256: digest next to each asset. Compare it with your file (PowerShell):

  Get-FileHash Voica.exe -Algorithm SHA256

  The hash must match the one shown on the release.

Code signing (which removes the warning) is planned once the project has enough public visibility to qualify for the free SignPath Foundation program.

First run

On first launch (with no key set) the Settings window opens. Paste your Groq API key, click Validate, then Save (it's encrypted with DPAPI). You can also set the key via the GROQ_API_KEY environment variable for development.

Usage

• Dictate: double‑tap Right Alt (default), speak, press Right Alt once to stop (Toggle mode). In PTT mode, hold to talk and release to send.
• While recording, the bar at the bottom of the screen shows the level: × cancels (the recording is discarded), ✓ stops and transcribes.
• The recognized text is inserted into the focused field and copied to the clipboard.
• Right‑click the tray icon for Settings, History, Check for Updates, and About.

With the dictation bar on (the default), the tray icon stays neutral. Turn it off and the icon reflects state instead: idle (blue), recording (pulsing red), transcribing (amber).

Settings

Setting	Notes
Dictation mode	PTT (hold) or Toggle
Hotkey	Preset single key or a custom combination
Double tap to start	Toggle mode only; on by default (0.35 s window)
Show a recording bar	The bottom‑of‑screen capsule; on by default
Output	Insert into field, or show a result window
Store audio recordings	On by default
Show a notification after inserting	The tray balloon; can be turned off
Check for updates on launch	Once a day, opt‑in
Delete audio older than	N days; 0 = keep forever
Vocabulary	Terms fixed by rule on both engines; also a cloud hint (last 800 chars used)
Fix terms by rules	On by default; the rules that need no key or network
Remove "uh", "um", "hmm"	On by default; off if you transcribe verbatim
Fix quotation marks	On by default; guillemets, pairing, space after a colon
Fix terms with AI	Optional Groq pass on top of the rules; off by default
Use the system proxy	On by default (Network tab); off = ignore Windows and go direct
Groq API key	Validate + Save (DPAPI); Show to reveal
Delete the local model…	Frees 214 MB, asks first; the cloud keeps working, and it can be downloaded again
Delete all data…	Wipes history, audio, key, settings (random‑phrase confirmation)

Data locations

Everything lives outside the executable in %APPDATA%\Voica\, so it survives updates:

• history.sqlite — transcription history
• audio\*.wav — stored recordings (16 kHz mono PCM)
• credentials.dat — DPAPI‑encrypted Groq key
• models\ — the local recognition model (if downloaded)
• settings.json — settings
• voica.log — local diagnostic log

Local offline engine

Switch Settings → General → Recognition engine to Local and Voica transcribes entirely on your PC with Sber's GigaAM v3 (MIT) — Russian speech with punctuation and text normalization, no network and no API key. The model (~215 MB, int8 ONNX) downloads once from this repo's model release with SHA‑256 verification, lives in %APPDATA%\Voica\models\, and can be deleted in Settings → Data (it asks first). The download can be cancelled while it runs, and nothing half‑downloaded is left behind. Notes:

• The vocabulary hint (§ Whisper prompt) works only with the cloud engine; AI term correction works with both (it needs a key and network).
• If cloud is selected but the network is down and the model is installed, Voica falls back to the local engine automatically and shows a small notice.
• With the local engine (and AI correction off), audio and text never leave your PC.
• ONNX conversion by istupakov/gigaam-v3-onnx (MIT).

Installing the model by hand

Some networks will not let 215 MB through, or will not let Voica out at all (see Behind a corporate proxy below). The model can be carried in on a USB stick instead: download the three files from the model release on any machine and put them, unrenamed, straight into %APPDATA%\Voica\models\:

File	Size	SHA-256
v3_e2e_ctc.int8.onnx	224 893 347	2e3fcb7a7b66030336fd10c2fcfb033bd1dc7e1bf238fe5cfd83b1d0cfc9d28e
v3_e2e_ctc.yaml	899	e67eca3a311ad7c8813d36dff6b8eeba7ad3459fd811d6faea2a26535754a358
v3_e2e_ctc_vocab.txt	2 007	142de7570b3de5b3035ce111a89c228e80e6085273731d944093ddf24fa539cd

Voica checks these checksums itself the first time it loads a model it did not download, and remembers the answer — so a truncated copy is reported as such instead of turning into gibberish recognition. To check them yourself first:

Get-FileHash "$env:APPDATA\Voica\models\v3_e2e_ctc.int8.onnx" -Algorithm SHA256

Behind a corporate proxy

Voica works where the only way out is a proxy. It uses the proxy Windows is configured with and authenticates as the signed-in user, so a domain proxy is answered over SSO — no password is typed into Voica or stored by it. The switch lives in Settings → Network, together with a line saying which route requests are actually taking. Turn it off to ignore the system settings and go straight out — a proxy left misconfigured in Windows blocks the app just as effectively as a missing one.

If the proxy refuses, every message names its address: that is the thing to hand to whoever administers it. If it will not be opened, the local engine plus the manual model install above gives you an app that needs no network at all.

Bring your own Groq key

Voica uses your own Groq API key (BYO-key) — the app never ships or shares anyone's key. Each user gets a free key at console.groq.com; usage is subject to Groq's Terms of Use. Free-tier limits (whisper-large-v3-turbo): 20 req/min, 2000/day, 7200 audio-seconds/hour — far more than dictation needs.

If you enable AI term correction (Settings → Vocabulary), Voica also calls a Groq chat model. The model isn't hardcoded: Voica reads the live model list for your key and picks the best available one (by default llama-3.3-70b-versatile, falling back through other families) — so when Groq retires or renames a model, the app heals itself instead of needing an update. You can also pin a specific model in the picker next to the toggle. If your Groq organization restricts model access, allow the chosen model at console.groq.com → Settings → Limits; otherwise the correction silently falls back to the raw transcription (fail-open by design).

Build from source

Requires the .NET 8 SDK.

# Build
dotnet build Voica.sln -c Debug

# Self-test (no GUI/network) — exit code 0 on success
Start-Process -FilePath "src\Voica\bin\Debug\net8.0-windows10.0.17763.0\Voica.exe" `
  -ArgumentList "--test-all" -Wait -PassThru -NoNewWindow

# Single-file self-contained release
dotnet publish src\Voica\Voica.csproj -c Release -p:PublishSingleFile=true

See CONTRIBUTING.md for architecture notes and the self‑test conventions.

License

MIT © Ivan Ushakov.