Typeless but Free

English · 中文

Local, private, hold‑to‑talk voice dictation with AI cleanup — for Windows. Like Typeless — but free & open source.

Hold a hotkey, talk, release. Your speech is transcribed on your own machine (faster‑whisper / CUDA), optionally cleaned up by an LLM (filler‑word removal, punctuation, light rewriting), and pasted straight into wherever your cursor is. A small floating card shows the transcription live as you speak.

Voice never leaves your machine for transcription. Only the (already‑transcribed) text is sent to the cleanup LLM — and you can turn cleanup off entirely for 100% local.

Hold ALT+V → speak → release → local Whisper → (optional) LLM cleanup
            → text auto‑pasted at your cursor

Features

• Local STT — faster-whisper on your GPU (CUDA) or CPU. No audio leaves your box.
• Live preview — see the transcription grow in real time while you talk.
• AI cleanup (optional) — any OpenAI‑compatible endpoint (DeepSeek, OpenAI, a local LLM…). Bring your own key, or turn it off.
• Auto‑insert — pastes into the focused window; no copy‑paste dance.
• Learns from your fixes — click the lingering card to correct a word; it remembers the fix (replacement rule + adds the term to hotwords) so it gets it right next time. 100% local.
• Configurable hotkey — hold‑to‑talk, double‑tap toggle, single key or chord (e.g. alt+v, ctrl_r, caps_lock).
• Pure Python + tkinter UI — dark, rounded, animated waveform that reacts to your mic volume.

Requirements

• Windows 10/11
• Python 3.10+
• (Recommended) an NVIDIA GPU for fast transcription. No GPU? It falls back to CPU automatically (slower).

Install

git clone <your-repo-url> voicetype
cd voicetype
powershell -ExecutionPolicy Bypass -File .\setup.ps1

setup.ps1 creates a venv, installs deps, and removes hf-xet (see Troubleshooting). First run downloads the Whisper model (large-v3-turbo ≈ 1.6 GB) once.

Configure

Copy the example and fill in your own values:

copy config.example.json config.json

config.json is git‑ignored — your key stays local.

To enable AI cleanup, set one of:

• llm_api_key in config.json, or
• env var LLM_API_KEY / DEEPSEEK_API_KEY / OPENAI_API_KEY, or
• llm_api_key_file pointing to a .env file containing one of those vars.

Want 100% local / no key? Set "cleanup_enabled": false. You’ll get the raw transcript inserted directly (keeps filler words, but nothing leaves your machine).

Config reference

key	default	meaning
hotkey	alt+v	trigger key/chord: alt+v, ctrl_r, caps_lock, f8, …
hotkey_mode	hold	hold = push‑to‑talk; double_tap = tap to start / tap to stop
whisper_model	large-v3-turbo	large-v3 (most accurate) · large-v3-turbo (≈2× faster, ~same) · medium/small/tiny (faster, less accurate)
whisper_device	auto	auto / cuda / cpu (auto+cuda fall back to CPU if no GPU)
language	null	auto‑detect; set "zh" / "en" to lock
simplify_chinese	true	convert Traditional → Simplified on the Chinese path (Whisper sometimes drifts to Traditional); set false if you want Traditional
beam_size	1	1 fastest; raise to 5 for marginally better accuracy
cleanup_enabled	true	LLM cleanup on/off (off = raw transcript, no key needed)
confirm_before_insert	false	true = show an editable popup before inserting
streaming_partial	true	live transcription while recording
linger_ms	4000	how long the bottom‑right card stays after inserting — click it within this window to fix/teach a word
max_display_lines	4	cap the card to N lines (older text scrolls off the top)
vocabulary	[]	terms to recognize more reliably, e.g. ["CLAUDE.md","agent"] (also editable in the settings UI)
insert_suffix	""	text appended after each insert (" " space / "\n" newline; avoid newline in chat apps)
restore_clipboard	true	restore your previous clipboard after inserting
max_record_seconds	120	auto‑stop recording after N seconds (0 = no limit)
paste_delay_ms	120	focus/paste delay; raise if a slow app drops the paste
llm_timeout_seconds	30	cleanup request timeout
llm_base_url	https://api.deepseek.com	any OpenAI‑compatible base URL
llm_model	deepseek-v4-flash	model name on that endpoint
hf_endpoint	https://huggingface.co	model download host — set to https://hf-mirror.com in mainland China
model_use_proxy	false	false = bypass system proxy when downloading models

Usage

.\run.bat

Hold Alt+V, speak, release. Watch the bottom‑right card. Done. Close the window or press Ctrl+C to quit.

Troubleshooting

Hard‑won notes (these cost a few hours to figure out — so you don't have to):

• Model config.json / tokenizer downloads as 0 bytes / "JSON parse error" → the hf-xet backend mangles small files on some networks/mirrors. setup.ps1 uninstalls it; if you reinstalled it, run pip uninstall -y hf-xet.
• cublas64_12.dll cannot be loaded → CUDA runtime is missing. Ensure nvidia-cublas-cu12, nvidia-cuda-runtime-cu12, nvidia-cudnn-cu12 are installed (they're in requirements.txt).
• Downloads reset / Connection aborted behind a VPN (Clash etc.) → the proxy is killing the connection. Keep model_use_proxy: false (default) so downloads bypass the proxy.
• In mainland China → set hf_endpoint to https://hf-mirror.com.
• Hotkey does nothing → run run.bat from your own desktop session (a global keyboard hook needs an interactive session).

How it works

1. Global hotkey listener (pynput) detects hold/release.
2. sounddevice records mic audio to memory; RMS drives the waveform.
3. faster-whisper transcribes — every ~0.6 s on the partial audio for the live preview, then a final pass on release. The previous utterance is fed as initial_prompt for continuity.
4. Optional cleanup via an OpenAI‑compatible chat endpoint.
5. Clipboard + simulated Ctrl+V inserts into the focused window.

Privacy

Audio is transcribed locally and never uploaded. With cleanup_enabled: false, nothing leaves your machine at all. With cleanup on, only the transcribed text is sent to your chosen LLM endpoint.

Credits

Built on faster-whisper, CTranslate2, pynput, and sounddevice.

License

MIT — see LICENSE.