TalkType AchiGever26

Use this command to install TalkType with WinGet:

winget install --id=achigever26.talktype -e

Free, open-source voice dictation for Windows. Hold a hotkey, speak, and clean text appears at your cursor. English, Hebrew, and 99 more languages — fully local

Free, open-source voice dictation for Windows. Hold a hotkey, speak, release — and clean, polished text appears at your cursor in any app.

Inspired by the best of OpenFlow and Wispr Flow but free forever, open source, and everything runs locally on your PC.

Features

Dictate anywhere — works in every app: email, Slack, Discord, VS Code, browsers, terminals
Hold or toggle — hold Ctrl+Win to talk (release to finish), or tap it to toggle; Esc cancels
Best-in-class engines, picked per language:
- English (+ 24 European languages): NVIDIA Parakeet TDT v3 — faster and more accurate than Whisper large, at 1/5 the size
- Hebrew (עברית): ivrit.ai fine-tuned Whisper — the best free Hebrew speech model, handles Hebrew-with-English naturally
- 99 other languages: OpenAI Whisper
Auto language detection — dictate Hebrew one moment, English the next; each dictation is routed to the best engine automatically
Quick language switch — press Ctrl+Alt+L (configurable) to cycle Auto → English → Hebrew → …; the choice flashes on screen, shows as a badge on the tray icon, and is also in the tray's right-click Language menu
AI cleanup (built-in, offline, on by default) — removes filler words ("um", "uh", "כאילו", "אה"), fixes punctuation and capitalization
Context-aware formatting — the same speech comes out right for where you're dictating: casual in chat apps (Slack, Discord, WhatsApp), full sentences in email, left verbatim in terminals (no capitalization "fixes" on your shell commands) and in AI assistants (Claude, ChatGPT, Perplexity, Gemini). Even browser tabs are detected: Gmail gets email style, WhatsApp Web gets chat style
AI rewrite (optional, opt-in) — full LLM-powered rewriting of messy speech:
- Ollama — runs locally on your PC, nothing leaves your machine (requires Ollama)
- OpenAI / Anthropic / Groq — bring your own API key. ⚠️ These send your dictated text (never audio) to the provider's servers; the settings screen warns you before you turn it on
- Planned: Windows Phi Silica (the built-in on-device rewrite model on Copilot+ PCs) once Microsoft's GPU/CPU rollout stabilizes
Custom dictionary — teach it names and jargon ("pie side six" → "PySide6")
Voice snippets — say "my sign off" and your saved template is inserted; supports {rest}, {date}, {time}, {clipboard}
Searchable history — full-text search over past dictations, stored locally in SQLite
No account. No telemetry. No subscription.

Privacy

Your audio never leaves your machine. Models run locally and work fully offline after the one-time download. Cleanup is rule-based and offline by default. The only feature that can ever send anything to the internet is the optional AI rewrite when set to a cloud provider (off by default, uses your own API key, and the settings screen shows a clear warning) — choosing Ollama keeps even that fully local. Full policy: PRIVACY.md.

Install

Via winget (recommended — no SmartScreen warning):

winget install achigever26.talktype

(The full ID is needed because another product also happens to be named TalkType.)

Or manually:

Download TalkType-Setup.exe from the latest release
Run it (no admin rights needed)

Then the first-run wizard downloads the speech models for your languages (145 MB – 1.6 GB depending on choices). Hold Ctrl+Win and start talking.

> SmartScreen note: Windows may warn about an unknown publisher — TalkType is a new open-source project without a paid code-signing certificate. Click "More info" → "Run anyway", or build from source below. Every release is built publicly by GitHub Actions from this source, with SHA256 checksums attached.

A portable ZIP (no installer) is also available on the releases page.

Code signing

TalkType binaries are not yet code-signed — that's why direct downloads show the SmartScreen warning (installing via winget avoids it). We plan to sign releases through the SignPath Foundation's free open-source program once the project meets their visibility criteria. Every release is built publicly by GitHub Actions from this source with SHA256 checksums, so you can always verify what you're running.

Requirements

Windows 10/11, 64-bit
4 GB+ RAM (8 GB recommended for the Hebrew model)
No GPU needed — everything runs on CPU. (CUDA GPU is supported for extra speed: install CUDA 12 + cuDNN 9 and pick "cuda" in settings.)

Build from source

git clone https://github.com/AchiGever26/talktype
cd talktype
python -m venv .venv
.venv\Scripts\pip install -e .[dev]
.venv\Scripts\python -m talktype

Run tests with pytest, lint with ruff check src tests.

How it works

Stage	Technology
Global hotkeys	pynput low-level keyboard hook
Audio capture	sounddevice (WASAPI), 16 kHz mono
Speech-to-text	Parakeet TDT v3 (ONNX int8) / faster-whisper (CTranslate2 int8)
Language routing	Whisper language-ID pass per dictation
Cleanup	Rule-based (offline) + optional LLM via your API key
Text insertion	Clipboard-paste that snapshots and restores your clipboard (all formats, incl. images), keeps its own writes out of Win+V clipboard history, RTL-safe. Alternative: per-character Unicode typing that never touches the clipboard
UI	PySide6 tray app

Known limitations

Text can't be injected into elevated (admin) windows — Windows blocks this by design. The text is copied to the clipboard instead.
Language auto-detection is per-dictation, not mid-sentence (mixed sentences like Hebrew-with-English-terms are handled fine by the models themselves).
The first dictation after startup takes a few extra seconds while the model loads.

License

MIT. PySide6 is used under LGPLv3 (dynamically linked). Speech models retain their upstream licenses: Parakeet (CC-BY-4.0), Whisper (MIT), ivrit.ai models (Apache-2.0 with attribution).