Windows push-to-talk dictation. Hold a hotkey, speak, release — the transcript is pasted into the focused window.
Windows-only. The paste backend calls Win32 keyboard APIs directly; the provider, config, audio, and VAD layers are platform-neutral. macOS and Linux paste backends are not implemented.
Python 3.10+.
git clone https://github.com/lfzds4399-cpu/voice2ai.git
cd voice2ai
python -m pip install -r requirements.txt
python app.pyFirst run opens a setup wizard: pick a provider, paste an API key, test, save.
Alternatively, run install.bat then start.bat. A PyInstaller spec in
build_tools/ produces a frozen .exe; the binary is unsigned, so SmartScreen
will warn on first launch.
Default hotkey: ctrl+shift+space. Built-in presets: f8, f9, right ctrl,
ctrl+alt+v. Press f9 to toggle continuous mode, which uses EnergyVAD to
detect pauses and transcribe the preceding utterance.
On hotkey press, voice2ai snapshots the foreground HWND and restores it before
the paste fires, preventing text from landing in the wrong application after a
mid-utterance focus change. Known terminals and editors receive
ctrl+shift+v; other windows receive ctrl+v.
| Provider | Model | Notes |
|---|---|---|
| SiliconFlow | FunAudioLLM/SenseVoiceSmall |
Mandarin-capable, reachable from mainland China. |
| OpenAI | whisper-1, gpt-4o-mini-transcribe |
Paid API, global. |
| Groq | whisper-large-v3-turbo |
Whisper-compatible endpoint, low latency. |
| Azure | whisper deployment |
Requires an Azure OpenAI resource. |
config.env accepts the canonical keys and the provider aliases listed in
.env.example (GROQ_MODEL, OPENAI_MODEL, SILICONFLOW_MODEL,
AZURE_SPEECH_KEY, AZURE_SPEECH_REGION).
voice2ai does not simulate keystrokes. It copies the transcript to the clipboard, releases held modifier keys, waits a few milliseconds, restores the target window if possible, then sends the paste shortcut.
If shift remains physically held when paste fires, Windows interprets
ctrl+v as ctrl+shift+v, which some terminals intercept and discard. The
modifier-release path is covered by tests/test_paste.py.
app.py entry-point shim
src/voice2ai/
main.py app orchestration
config.py settings load/save
audio.py microphone capture and pre-roll
hotkey.py global hotkey listeners
paste.py Win32 clipboard paste backend
vad.py continuous-mode VAD
diagnostics.py dependency / mic / network / provider checks
providers/, ui/ STT providers and Tk UI components
tests/ offline unit tests
build_tools/ PyInstaller spec
python -m pip install -r requirements.txt
python -m pip install pytest ruff
python -m pytest tests/ -q
python -m ruff check src tests app.pyTests run offline. Provider APIs, real keyboard input, and GUI automation are not exercised.
Audio is sent only to the configured STT provider. No telemetry, analytics,
auto-update, or listening sockets. API keys are stored in config.env, which
is gitignored. Security reports: SECURITY.md.
MIT. See LICENSE.