OS Voice

Your voice is your new keyboard.

Local-first, AI-powered voice-to-text that types into any app — private by design.

A product of OSA · by Miosa.ai

---

What it is

OS Voice turns your voice into clean, structured text in any application. Press a hotkey, talk, and your words are transcribed locally, optionally polished by a local AI, and typed straight into whatever you're using — your editor, browser, chat, anywhere. It can run 100% offline: no account, no cloud, no telemetry.

---

✨ Features

🎙️ Dictate anywhere — press-to-talk or hands-free

• Tap Ctrl+Space to lock the mic and dictate a whole paragraph without holding anything; tap again to stop.
• Hold the key for classic push-to-talk.
• Cancel a take instantly with the pill's ✕ or the cancel hotkey.
• Output is typed directly into the focused app (uses ydotool/wtype on Wayland, native input elsewhere).

⚡ Local Whisper transcription, GPU-accelerated

• Runs whisper.cpp on your machine — CPU or GPU (Vulkan / CUDA).
• Model sizes from Tiny → Large-v3-Turbo; pick speed vs accuracy.
• On a modern GPU the Turbo model transcribes dozens of times faster than real time (a 10-second clip in a fraction of a second).
• Fully offline — your audio never leaves the device.

🧠 Local AI cleanup — the smart layer

Pipe the raw transcript through a local LLM (via Ollama) to get polished, structured writing:

• Context-aware commands — say "period" and it adds a . when you mean punctuation, but keeps the word when you mean the thing. Say "quote … unquote" and it wraps the phrase in quotation marks.
• Filler removal — drops "um", "uh", "you know", false starts, and repetitions.
• Topic-based paragraphs — automatically starts a new paragraph when you switch subjects, so a rambling brain-dump comes out structured.
• Faithful — keeps your words and voice; it cleans, it doesn't invent content.
• No em-dashes (and other house-style rules) — fully configurable via tones.
• Swap the cleanup model anytime (qwen3.5, qwen2.5, llama3.2, granite4.1, gemma4, …).

✍️ Writing styles (tones)

Switch how your words come out — on the fly with the arrow keys:

• Polished — natural, cleaned-up prose
• Structured — topic paragraphs, lists, tidy formatting
• Email — greeting / body / sign-off
• Verbatim — exactly what you said, no editing

📖 Personal dictionary

Teach OS Voice your world so it's always right:

• Glossary terms (names, brands, jargon) bias recognition.
• Replacement rules ("when I say X, write Y") fix things instantly.

🌍 More

100+ languages with hotkey switching · searchable transcript history · everything stored locally in SQLite · encrypted local key storage.

---

🔧 How it works

  🎤 mic
   │
   ▼
 Whisper engine (local, GPU)  ──►  raw transcript
   │
   ▼
 Local LLM cleanup (Ollama, optional)  ──►  polished + structured text
   │
   ▼
 ⌨️  typed into your active app

"Rust is the API, TypeScript is the brain" — all logic lives in TypeScript; Rust provides native capabilities (audio, input injection, the Whisper sidecar, the overlay pill).

Stack: Tauri 2 (Rust + TypeScript/React) · whisper.cpp (CPU + Vulkan/CUDA) via a sidecar engine · Ollama for AI post-processing · SQLite · Turborepo monorepo.

---

🎚️ Recommended models

Transcription (Whisper) — set in Settings → Transcription:

Model	Best for
large-v3-turbo (Turbo)	GPU — top accuracy, many× faster than real time
small	CPU — great balance of speed + accuracy
base / tiny	low-end hardware / maximum speed

AI cleanup (Ollama) — set in Settings → AI Post-Processing. Pick by your hardware:

Model	Notes
gemma4:12b	Recommended — best topic-paragraph structure, faithful to your words (~1s on a strong GPU)
qwen2.5:14b	Solid all-rounder — faithful, good structure
qwen3.5:9b	Fastest faithful option; lighter on paragraph structure
granite4.1:8b	Faithful, instruction-tuned (IBM)
llama3.2:3b	Tiny + near-instant (~0.3s) — great for modest hardware

All run locally via Ollama. Heavier models give better structure at the cost of more VRAM/latency; smaller ones are instant. The cleanup is faithful by design — it reformats your words and never invents content. Requires a recent Ollama (ollama --version ≥ 0.30) to pull the newest models.

💻 Hardware requirements

OS Voice runs entirely on your own machine. Whisper transcription always runs locally; AI cleanup (Ollama) is optional and also fully local. Sizing depends on which Whisper model and which Ollama model you pick — see Recommended models above.

Available Whisper models (download automatically on first use, from huggingface.co/ggerganov/whisper.cpp): tiny, base, small, medium, large-v3 (large), large-v3-turbo (large-turbo — the recommended Turbo model).

Tier	Hardware	Whisper model	AI cleanup (Ollama)	Notes
Low-end	Any modern CPU	tiny / base	Off	Smallest, fastest-to-download models; slower transcription, fine for short bursts
Mid	CPU, or entry-level GPU	small (CPU) / medium (GPU)	Optional — light model (llama3.2:3b, qwen3.5:9b)	Good balance of speed and accuracy
High	Dedicated GPU (Vulkan/CUDA) or Apple Silicon	large-v3-turbo (Turbo)	Yes — gemma4:12b / qwen2.5:14b or bigger	Turbo needs GPU acceleration (Vulkan on Linux/Windows, CUDA, or Apple Silicon) to hit its "many× faster than real time" speed — on CPU-only it's much slower, like the other large models
Reference machine	Mac Studio M3 Ultra, 96GB unified memory	large-v3-turbo	Yes — large local or Ollama cloud models	Enough headroom to run the top Whisper model and a large cleanup model at the same time with no compromises

Approximate guidance (exact figures vary by build/quantization — treat as ballpark, not a spec):

• Whisper models: tiny/base are small downloads (tens of MB) and run comfortably on CPU; small/medium are roughly in the hundreds of MB up to ~1.5GB; large-v3/large-v3-turbo are roughly 1.5–3GB and benefit heavily from GPU acceleration. Each model is downloaded once and cached locally.
• Ollama cleanup models: a 3B model like llama3.2:3b needs only a few GB of RAM/VRAM and responds almost instantly; gemma4:12b/qwen2.5:14b-class models need roughly 8–16GB of RAM/VRAM depending on quantization; bigger models need more. Apple Silicon uses unified memory (shared between CPU/GPU), so a high-RAM Apple Silicon Mac can run both Turbo transcription and a large local LLM at once.
• Disk: budget a few GB if you try multiple Whisper models, plus a few GB per Ollama model you pull.

📥 Installing (end users)

Prebuilt installers are published to GitHub Releases. Pick the package for your OS.

Linux — .deb, .rpm, or a portable AppImage. The .deb/.rpm install these runtime dependencies automatically (install them manually if you use the AppImage):

Purpose	Debian/Ubuntu (.deb)	Fedora/openSUSE (.rpm)
Input injection (X11)	libxdo3, xdotool	xdotool
GPU transcription (Vulkan)	libvulkan1	vulkan-loader
Overlay pill (Wayland layer-shell)	libgtk-layer-shell0	gtk-layer-shell
Wayland typing	wtype	wtype

Windows — .msi installer (bundles the WebView2 bootstrapper), or the NSIS .exe installer; a portable, no-admin .exe installer is also published.

macOS — .dmg (drag-to-Applications, universal Apple Silicon + Intel binary) or .pkg. Requires macOS 13.3 (Ventura) or later.

After installing:

• Whisper transcription models download automatically the first time you select them in Settings → Transcription — no manual setup.
• AI cleanup is optional and requires Ollama installed separately. Install Ollama, then pull a model (see Recommended models above), e.g.:

  ollama pull gemma4:12b

  Then enable it in Settings → AI Post-Processing.

🚀 Build & run (Linux)

Requires Node 18+, Rust, and the system deps in apps/desktop/scripts/setup-linux.sh (plus glslc for the GPU build, and ydotool/wtype for Wayland typing).

pnpm install
cd apps/desktop
pnpm run prepare:sidecars          # builds the Whisper engine + native pill
pnpm exec tauri build --bundles deb

For local AI cleanup, install Ollama and pull a model (e.g. ollama pull gemma4:12b — see Recommended models above), then enable it in Settings → AI Post-Processing.

See CLAUDE.md and apps/desktop/scripts/setup-linux.sh for full platform details.

---

🙏 Credits & license

OS Voice is a rebrand and continuation of the excellent open-source Voquill by Voquill, Inc. and its contributors. The architecture, Whisper engine, overlay, and essentially all of the application code originate from Voquill — this project would not exist without their work. Please go support them. ❤️ (See CREDITS.md.)

Licensed under the GNU AGPLv3 (see LICENSE), the same license as upstream Voquill — the full source is here and all modifications remain open.

"OS Voice", "OSA", and "Miosa" branding © Miosa.ai. The underlying application is © 2025–present Voquill, Inc., used and modified under the AGPLv3. Bundled config files use placeholder/redacted credentials — supply your own to enable optional cloud features.