Local, private, GPU-accelerated voice-to-text dictation for Windows
Speak Slovenian or English by default — or pin any Whisper language.
Local by default. Optional self-hosted recognition server and text-only AI cleanup.

---

What it does

Hold F1, talk, release. Short transcription usually lands in the focused window in ~0.5–1 second — Notepad, Discord, browser chat, IDE, game chat. Long recordings pre-transcribe complete speech blocks while you talk, retain the full microphone stream for automatic recovery, and still paste exactly once after release. The default is fast English + Slovenian Auto, with dedicated standard and slang Slovenian modes. Users can also search and pin any of Whisper large-v3's 100 languages without downloading another model. Completed text can be kept in a bounded local recovery history, so a failed target application never costs the whole prompt.

Optionally, VoicePrompt can remove speech clutter, fix spoken grammar, or turn a rough transcript into a cleaner AI prompt before delivery. This is disabled by default and never changes the local audio pipeline.

Advanced users can explicitly switch recognition to a self-hosted OpenAI-compatible speech server. Local remains the default; the UI clearly warns when recorded audio would leave the PC and whether the connection is encrypted.

Measured on an RTX 5080:

Model	VRAM while idle	Load time	Per 8s utterance	Slovenian quality
large-v3-turbo (fast)	~2.2 GB	~3 s	~0.35 s	okay
large-v3 (accuracy)	~4 GB	~5 s	~0.6 s	best

Both run entirely locally on your GPU through faster-whisper (CTranslate2). The driving app is the open-source faster-whisper-dictation daemon.

Architecture

 F1 hold
    |
    v
+----------------------+   +----------------------+   +----------------------+
| Hotkey listener      |-->| Windows audio        |-->| faster-whisper       |
| Win32 native hotkey  |   | 16 kHz mono          |   | large-v3 CUDA fp16   |
| selected key consumed|   | live level meter     |   | language selection   |
+----------------------+   +----------------------+   +----------+-----------+
                                                                 |
                                                                 | transcript
                                                                 v
+----------------------+   +----------------------+   +----------------------+
| Focused application  |<--| Clipboard paste      |<--| AI cleanup (optional)|
| Chat, IDE, browser   |   | Ctrl+V, then restore |   | raw, grammar, prompt |
+----------------------+   +----------------------+   +----------------------+

• Hold mode: recording only while you hold the key → no accidental captures.
• VAD filtering: silence/speech detection chops dead air before decoding.
• Personal vocabulary: optional prompt and hotword fields can bias decoding toward names and exact technical terms without favoring one language by default.
• Personal corrections: explicit misheard => intended rules fix recurring names and terms locally before optional AI cleanup.
• Daemonized: runs headless via pythonw, survives reboot via the Startup shortcut.

Download and install

Requirements: 64-bit Windows 11, Python 3.11+, an NVIDIA GPU with a current driver, and roughly 10 GB of free disk space for the runtime and model.

1. Open the latest VoicePrompt release and download VoicePrompt-v1.22.0-windows-x64.zip.
2. Extract the ZIP, open PowerShell in that folder, and run:

powershell -ExecutionPolicy Bypass -File .\install.ps1

The installer creates the private Python environment, installs the exact pinned speech runtime, applies the Windows integration fixes, installs the self-contained tray app, and creates desktop and Start Menu shortcuts. Open VoicePrompt, then hold F1 to talk. The first start downloads the selected model (~3.1 GB for large-v3) once.

Future releases can be installed from System → Application updates: click Check now, then Download & install. VoicePrompt accepts only the matching official GitHub package, verifies its size, GitHub asset digest, published SHA-256 checksum, safe archive paths, required files, and version, then runs the same settings-preserving installer. Checks and installs remain manual—there is no background request or silent update.

The executable is not code-signed yet, so Windows may show an “unknown publisher” warning. Only run packages downloaded from this repository's official Releases page. The SHA-256 checksums are attached to every release.

Build from source

# 1. Create the venv and install the tested runtime (requires Python 3.11+ and an NVIDIA GPU)
py -m venv "$env:USERPROFILE\.voice-typing\venv"
& "$env:USERPROFILE\.voice-typing\venv\Scripts\pip" install --only-binary=:all: -r requirements.txt

# 2. Apply the Windows fixes (see "Patches")
powershell -ExecutionPolicy Bypass -File scripts\apply_patches.ps1

# 3. Build the self-contained Windows x64 tray UI (requires .NET SDK 10+)
powershell -ExecutionPolicy Bypass -File scripts\build_ui.ps1

# 4. Start VoicePrompt. It runs in the system tray and starts the daemon.
.\ui\publish\VoicePromptTray.exe --tray

First start downloads the model (~1.6 GB turbo / ~3.1 GB full) into ~/.cache/huggingface/hub — one-time.

🖥️ Tray UI (ui/VoicePromptTray)

A responsive dark Windows tray app (C# / .NET 10 WinForms) that manages the whole setup:

• Setup overview — confirms runtime, hotkey, microphone, and recognition readiness at a glance, with direct recovery and support actions.
• Three interface themes — Graphite, Evergreen, and Ember apply instantly, stay local, and share the same restrained dark visual system.
• Native dark window — the Windows caption, frame, controls, and scroll surfaces match the app instead of flashing light chrome around it.
• Focused workspaces — Dictation, Audio, Engine & AI, Recovery, and System pages keep everyday setup simple while leaving expert controls available.
• System tray — runs minimized next to the clock; double-click the icon (or the desktop VoicePrompt shortcut) to open settings. The tray menu can instantly copy the latest saved transcript, start/stop/restart the daemon, and quit the app.
• Recording overlay — choose a compact live waveform, equalizer bars, or a reactive microphone orb. The clean graphite popup follows the active screen, appears immediately while the hotkey is held, and never takes keyboard focus.
• Live input test — the Audio page shows quiet, good-signal, and very-loud microphone levels while the hotkey is held by reusing the overlay stream, with no second capture or extra GPU work.
• Hotkey recorder — click the box, press one key (F1, Space, 7…) or a combo (Ctrl+Shift+F1, Alt+Space…), Enter confirms, Esc cancels. Supports hold (press & hold to talk) or toggle modes.
• Flexible output — paste directly into the focused app by default, or use Copy only when a target blocks synthetic paste and place the completed transcript manually.
• Exact voice commands — optionally speak a complete English or Slovenian command for a new line, new paragraph, bullet, undo, or cancel; normal sentences never trigger commands by substring.
• Reusable snippets — save up to 50 local text templates and insert one by its exact English or Slovenian spoken name, including multi-line content without AI or network delay.
• Writing modes — Verbatim stays fully local and instant; optional Clean, Grammar, and Prompt modes use a configured provider with a strict deadline, same-language instructions, and complete-original fallback.
• Application profiles — optionally override writing and output mode for an exact running-app executable; unmatched apps inherit global settings with no process lookup or added delay.
• Recognition location — keep the recommended local GPU engine, or explicitly use a self-hosted OpenAI-compatible transcription server with a no-audio health check and clear transport privacy status.
• Fast long recordings — pre-transcribes complete speech blocks on the existing model worker, preserves the full recording for recovery, and produces one ordered paste after release.
• Recovery — keeps a configurable 5–100 recent transcripts locally, compares the delivered result with the untouched original, and lets you copy either version. Audio is never stored.
• Personal corrections — applies approved phrase replacements deterministically with no model call or added network delay.
• Portable language profiles — exports or imports language, recognition context, hotwords, and personal corrections in a small JSON file. Profiles never contain API keys, transcript history, hotkeys, microphone names, or hardware settings.
• Settings backup — exports and review-first imports portable dictation, recognition, audio-detection, writing, recovery, correction, snippet, and application-profile settings without API keys, transcripts, microphone identity, startup state, or machine paths.
• All settings — language (English + Slovenian Auto, dedicated defaults, or any pinned Whisper language), decoding prompt, VAD threshold & timing, microphone (enumerated live) and sample rate, model (large-v3 / large-v3-turbo), compute type, GPU/CPU, temperature, hotwords.
• Guided recovery — tested recognition defaults, live microphone refresh, Windows Sound Settings, privacy-safe performance statistics and copied diagnostics, and one-click log/config access.
• Verified in-app updates — manually checks GitHub with a three-second timeout, then offers a confirmed Download & install action. Stable is the default; Preview is explicit opt-in. Downloads are checked against GitHub's digest and the release checksum before the settings-preserving installer runs; nothing checks, downloads, or installs silently.
• Keyboard and accessibility — predictable tab navigation, screen-reader names, Ctrl+S to save, Esc to hide, and Ctrl+1 through Ctrl+6 for page navigation.
• Save & Restart writes the live config, keeps your comments, and restarts the daemon in one click.
• Start with Windows checkbox manages its own startup shortcut; the daemon auto-starts with the UI.

Rebuild after changes: scripts\build_ui.ps1 (outputs ui\publish\VoicePromptTray.exe).

⚙️ Configuration

The tray UI edits the live config — %LOCALAPPDATA%\faster-whisper-dictation\faster-whisper-dictation\config.toml (snapshot in this repo: config.toml). Settings load at daemon start — the UI restarts it for you on Save.

Key	Meaning	Example
[hotkey] binding	Global hotkey (single key or combo)	"f1", "ctrl+shift+f1", "alt+space"
[hotkey] mode	hold = press & hold; toggle = press once	"hold"
[server] url	Compatible transcription server base URL; used only when engine type is server	"http://localhost:8000"
[server] timeout	Maximum server wait after releasing the hotkey	60
[server] model	Whisper model	"Systran/faster-whisper-large-v3"
[server] language	"" = fast English/Slovenian Auto; otherwise pins one supported language	"sl-slang", "sl", "en", "de", "ja"
[server] prompt	Optional personal names / exact vocabulary bias	empty (language-neutral)
[voiceprompt] slovenian_slang	Saves the visible colloquial vocabulary profile	true / false
[voiceprompt] buffered_transcription	Pre-transcribes long speech blocks; full-audio fallback and one final paste	true
[voiceprompt] output_mode	"paste" types into the focused app; "clipboard" sends no paste shortcut	"paste"
[voiceprompt] voice_commands	Enables exact whole-utterance English and Slovenian commands	false
[vad] threshold	Speech sensitivity (0–1)	0.6
[vad] max_speech_s	Internal VAD segment size; not a recording cutoff	180.0
[engine] type	Private local engine or explicit compatible server	"local", "server"
[engine] compute_type	float16 GPU / int8 CPU	"float16"
[audio] device	"" = Windows default input (HyperX Quadcast)	""

English + Slovenian Auto

Auto is intentionally optimized for the two languages this app targets. The primary pass stays language-neutral, so English cannot inherit Slovenian slang examples and Slovenian cannot inherit English instructions. Confident English and Slovenian remain a single-pass path, keeping normal dictation fast.

A detected English result is never replaced by Slovenian. A moderate-confidence Slovenian guess that contradicts recent English or competing language evidence is cross-checked once as English, and the retry replaces it only when the combined evidence wins. If prepared blocks from one long recording disagree, VoicePrompt retries the retained complete recording instead of pasting mixed languages. Finnish, Spanish, Latin, and other unrelated Auto guesses remain constrained to one bounded English or Slovenian recovery pass. Slovenian recovery receives the compact colloquial profile for forms such as dej, lohk, kva, tko, tle, zdej, pol, ful, and štima.

Additional languages

The Dictation page includes a searchable catalog of all 100 language codes supported by the installed Whisper large-v3 model. Choosing a result switches from bilingual Auto to a pinned recognition profile. Pinning bypasses language detection, keeps transcription in the spoken language, and does not invoke Whisper's separate translation task. The language profile is only a small setting: the existing multilingual model is reused, so there is no extra model download, network call, or GPU memory cost. Returning to Auto restores the tested English + Slovenian behavior.

Compatible recognition server

Under Engine & AI → Recognition engine, choose Compatible server, enter the server base URL, set a bounded wait, and click Test server. The test performs only GET /health; it sends no audio, transcript, credentials, or request body. Actual dictation sends one completed WAV recording to POST /v1/audio/transcriptions, following the upstream OpenAI-compatible REST contract. Local processor, precision, and background pre-transcription controls are disabled because inference occurs on the server.

The default http://localhost:8000 keeps audio on the same PC. Remote HTTPS is supported with a clear notice that completed recordings leave the computer. Remote unencrypted HTTP is allowed for deliberate trusted-network setups but receives a prominent warning. VoicePrompt does not add an authorization header, so use a server that is already protected by its own trusted network or reverse proxy. WhisperLiveKit is the upstream-recommended compatible server; the base daemon documentation covers its server mode.

Optional AI cleanup

The AI text cleanup card supports four writing modes:

• Verbatim: the default. No request, no per-utterance network or file access, and no added delay.
• Clean: removes obvious filler words and immediate repetitions, then repairs punctuation and capitalization without changing grammar, phrasing, tone, language, or concrete details.
• Grammar: conservatively fixes punctuation, capitalization, and obvious grammar without translating, paraphrasing, or removing wording.
• Prompt: turns rough speech into a concise, structured AI prompt while preserving requirements, names, code, numbers, and URLs.

The endpoint can be any OpenAI-compatible /v1/chat/completions service. For a private local setup, install Ollama for Windows, run ollama pull qwen2.5:3b, and keep the default endpoint and model. Ollama documents its OpenAI-compatible endpoint here. A cloud provider also works by entering its endpoint, model, and API key.

Only the completed transcript text is sent when cleanup is enabled; microphone audio stays local. Saved API keys are encrypted for the current Windows account. Live requests reuse one connection, cap the response size, wait at most 400–3000 ms (900 ms by default), and paste the untouched local transcript after any timeout or invalid response. The Test button can take up to six seconds because it also wakes a sleeping local model; this longer test allowance does not apply to live dictation.

Optional voice commands

Enable Voice commands on the Dictation page. A command runs only when the entire recognized utterance matches one phrase (terminal punctuation and quotation marks are ignored), so phrases inside ordinary dictation stay ordinary text. Commands run locally before AI cleanup and are not added to Recovery.

Action	English	Slovenian
New line	New line	Nova vrstica
New paragraph	New paragraph	Nov odstavek
Bullet	Bullet point	Alineja
Undo in focused app	Undo	Razveljavi
Discard this transcript	Cancel	Prekliči or Preklici
Insert a saved snippet	Insert snippet name	Vstavi predlogo name

Snippets are edited under Dictation → Reusable text, one per line as name => content. Use \n where the inserted text should contain a line break. Snippets remain local, are loaded once when the runtime starts, and share the same exact-match safety gate as the built-in commands.

Settings and vocabulary backup

Use System → Data portability to export or import one validated JSON backup. Import is review-first: it fills the settings pages but changes neither the running daemon nor saved files until Save & restart is clicked. The backup includes the global hotkey, language and vocabulary, output behavior, recognition engine/server settings and VAD tuning, AI mode/provider settings, recovery preferences, corrections, snippets, and application profiles.

The encrypted API key, transcript history, microphone identity, Windows startup state, window preferences, logs, and machine-specific paths are never exported. Endpoint URLs containing embedded credentials, query strings, or fragments are rejected to prevent accidental token leakage.

Application profiles

Use Engine & AI → Application profiles to select a running app and add an exact executable rule. Each line uses app.exe => writing, output. Writing can be inherit, verbatim, clean, grammar, or prompt; output can be inherit, paste, or clipboard. Matching is case-insensitive but otherwise exact—no paths, wildcards, background monitoring, or hidden changes. Profiles are empty by default and unmatched applications keep the global behavior.

🔧 Patches (required on Windows)

VoicePrompt's tested Windows integration layer is applied after every reinstall or upgrade:

1. cli.py — _pid_alive() used os.kill(pid, 0), which raises OSError (WinError 87) on Windows and broke status / stop. Now uses OpenProcess via ctypes.
2. typer.py — clipboard calls had no argtypes/restype, so 64-bit HANDLEs were truncated to 32 bits → access violations when pasting. All Win32 calls now declare their signatures, retry bounded clipboard contention, verify the full Unicode payload, and fail visibly instead of dropping text.
3. engine/local.py / slang_retry.py — runs a language-neutral English/Slovenian Auto pass, skips retries for strong English transcripts, and constrains unrelated-language mistakes to the supported bilingual route.
4. engine/local.py — logs detected language + confidence per utterance (auto-detect diagnostics).
5. engine/local.py — passes prompt, temperature, hotwords, and VAD controls to local faster-whisper. These settings otherwise have no effect in the upstream local engine.
6. engine/local.py / decoding_options.py — keeps beam-5 accuracy while bounding every language pass to one decode, with independent 30-second windows, repetition penalty, and native no-repeat protection.
7. daemon.py / meter.py — publishes recording state immediately on hotkey activation, then streams microphone levels and waveform samples through named shared memory without a second audio capture or disk polling.
8. config.py — hotkey validation accepts single keys (letters, digits, f1–f11, f13–f24, space, enter, …) and Ctrl/Alt/Shift combos while rejecting Windows-reserved bindings.
9. hotkey/listener.py / windows_hotkey.py — registers the shortcut through Windows' native global-hotkey message queue, so keys such as F1 do not also trigger browser help or application commands. Hold release is read from physical key state, key repeat is suppressed by the OS, callback failures are isolated, and the low-level hook that could silently stop after idle is not used on Windows.
10. typer.py / local text tools — applies approved corrections, stores bounded local recovery, and optionally cleans completed text before the clipboard is opened, with a strict deadline and raw-text fallback.
11. daemon.py — retains every held-recording audio chunk instead of silently discarding everything after 90 seconds; VAD segmentation remains bounded without limiting the complete recording.
12. daemon.py / buffered_transcription.py — serially pre-transcribes complete long-recording speech blocks without typing partial text, preserves block order, and retries the retained full audio after any empty, failed, or incomplete background result.
13. typer.py / output_mode.py — routes the final transcript exactly once to automatic paste or verified clipboard-only delivery, without emitting paste keystrokes in Copy-only mode.
14. typer.py / voice_commands.py / text_snippets.py — recognizes only enabled whole-utterance English/Slovenian commands and saved snippets, skips AI and history for them, and routes each result exactly once.
15. typer.py / app_profiles.py / ai_rewriter.py — resolves an optional exact focused-app rule once after transcription, overrides only writing/output behavior, and inherits global settings after any missing, inaccessible, unmatched, or invalid rule.
16. engine/server.py — maps the internal Slovenian slang profile to the standard sl API code and omits an empty language field so compatible servers can auto-detect safely.
17. audio.py — validates a saved input device on every recording and falls back to the current Windows default when that device was unplugged, renamed, or removed.

Run: powershell -ExecutionPolicy Bypass -File scripts\apply_patches.ps1

🧪 Testing

The E2E harness simulates what a human does (no spoken voice needed):

• tests/e2e_test.ps1 — opens a live text target window, presses the hotkey via keybd_event, plays an audio file through the speakers/microphone path, and proves the transcribed text lands there (timestamps the release→paste latency).
• tests/bench_one.py / tests/accuracy_metrics.py — model load time, VRAM, WER, CER, language accuracy, repetition rate, and p50/p95 latency from an optional reference manifest.
• tests/probe_devices.py — enumerates PortAudio input devices.
• tests/test_ai_rewriter.py — exercises both cleanup modes, warm connection reuse, API authentication, response guards, strict timeouts, and raw fallback against a local mock provider.
• tests/test_slang_retry.py — verifies language-neutral primary decoding, bilingual recovery, transcript confidence gates, and safeguards against translating real English.
• tests/test_decoding_options.py — verifies latency-bounded temperature handling and native repetition-loop protection on both English and Slovenian decoding passes.
• tests/test_local_text.py — verifies personal corrections and bounded, optional, Unicode-safe local recovery.
• tests/test_buffered_transcription.py — verifies speech-block ordering, one-result assembly, short-recording compatibility, and complete-audio fallback triggers.
• tests/test_output_mode.py — verifies safe defaults and proves Copy-only delivery cannot call the synthetic-paste route.
• tests/test_voice_commands.py — verifies default-off behavior, exact English/Slovenian recognition, Unicode output, and substring false-positive protection.
• tests/test_server_engine.py — verifies bounded OpenAI-compatible WAV requests, Slovenian/Auto language routing, and empty-result fallback on timeouts or malformed responses.
• tests/test_text_snippets.py — verifies bounded Unicode snippet loading, bilingual exact resolution, malformed-data fallback, and false-positive protection.
• tests/test_windows_hotkey.py — verifies shortcut parsing plus native Windows press/release delivery after idle and across repeated activations.
• tests/test_patch_migrations.ps1 — verifies clean, legacy, and immediately previous release upgrades compile and remain byte-for-byte idempotent when the patcher is reapplied.
• ui/LayoutCheck — verifies every settings layout plus cold overlay activation at full opacity.
• ui/ConfigManager.Tests — verifies the comment-preserving config editor plus privacy-safe, validated settings/vocabulary backup round trips (run: dotnet run --project ui\ConfigManager.Tests).

Verified end-to-end results (simulated): a clear English utterance landed 0.78 s after key release; an ambiguous Slovenian utterance requiring the safety decode landed in 1.86 s. The original batch path retained a 129.9-second recording's opening sentence, all 18 checkpoints, and unique final sentence. With fast long recordings enabled, a real-time 99.2-second English sample produced one ordered paste 0.49 s after release and matched the full one-pass decode's measured 7.04% WER on the same reference.

🩹 Troubleshooting

Symptom	Cause / fix
Library cublas64_12.dll is not found	CUDA 12 DLLs not on PATH → rerun install.ps1 so the pinned nvidia-*cu12 packages are repaired; run_daemon.pyw prepends their bin dirs automatically
Nothing typed but recording starts	Transcription crashed → read %USERPROFILE%\.voice-typing\daemon.log; check language = "" (not "auto") and patches applied
Hotkey cannot be saved or registered	F12 and Windows-logo combinations are reserved by Windows; choose another key or use Ctrl, Alt, and Shift modifiers
Bad or slangy Slovenian accuracy	Leave language on Auto for English/Slovenian routing; Slovenian recovery adds colloquial vocabulary automatically. Keep large-v3 for best accuracy and add only personal names or exact terms to Prompt/Hotwords
Mic not captured	Windows Settings → Privacy → Microphone → allow desktop apps (and make sure the Quadcast is the default input)
AI test fails or times out	Confirm the endpoint and model, start the provider, then click Test again; live dictation will paste the original transcript whenever cleanup is unavailable

🖥️ Lifecycle

• Auto-start on login: shell:startup shortcut → VoicePromptTray.exe --tray (tray UI; starts the daemon itself)
• Shortcuts: one canonical VoicePrompt link on the desktop and in the Start Menu; upgrades remove only recognized VoicePrompt-owned legacy links.
• Logs: %USERPROFILE%\.voice-typing\daemon.log (rotates at 2 MB; three archives retained)
• State: daemon PID/status via faster-whisper-dictation.exe status
• UI prefs: %APPDATA%\VoicePrompt\prefs.json
• AI settings: %APPDATA%\VoicePrompt\ai.json (API key protected with Windows account encryption)
• Recovery history: %APPDATA%\VoicePrompt\history.json (bounded text only; optional; no audio)
• Personal corrections: %APPDATA%\VoicePrompt\corrections.json

📜 Credits

• faster-whisper-dictation — the dictation daemon (MIT)
• faster-whisper — CTranslate2 whisper runtime (MIT)
• OpenAI Whisper large-v3 — the model (MIT)
• Logo artwork: generated for VoicePrompt; multi-resolution Windows icon packaged by scripts/make_icon.ps1

Roadmap

• 1.21, daily workflow: guided microphone calibration, explicit correction learning from Recovery, more global actions, and shortcut-conflict guidance.
• 1.22, optional context: privacy-visible selected-text and focused-app formatting, kept off by default and blocked in secure fields.
• 1.23, broader hardware: benchmark-gated CPU, AMD, and Intel engine options plus an explicit VRAM-saving mode.
• Later: file transcription, scratchpad reprocessing, speaker separation, and a signed installer. Code signing requires a trusted signing certificate and stays separate from unsigned development builds.