Skip to content

README_en.md

Latest commit

 

History

History
executable file
·
160 lines (118 loc) · 6.58 KB

README_en.md

File metadata and controls

executable file
·
160 lines (118 loc) · 6.58 KB

SenseType

Local voice typing for Windows, powered by SenseVoice -- 15x faster than Whisper, with better Chinese accuracy.

中文

Why SenseType?

  • Type anywhere -- press a hotkey, speak, and text appears at your cursor in any Windows application (browser, Word, VS Code, WeChat, etc.)
  • Simple and lightweight -- a single Python package, no complex setup, no background services
  • Free and open source -- no subscriptions, no API keys, no usage limits
  • Best-in-class Chinese -- powered by SenseVoice-Small, which achieves 15-20% lower error rate than Whisper on Chinese
  • Blazing fast -- non-autoregressive architecture processes 10s of audio in ~70ms, 15x faster than Whisper-Large
  • Fully local -- all processing happens on your machine, nothing is uploaded
  • Just works -- auto-detects GPU/CPU, built-in punctuation and voice activity detection, no post-processing needed

SenseType vs Whisper-based tools

SenseType Whisper-based tools
Model SenseVoice-Small (Alibaba) Whisper / faster-whisper
Architecture Non-autoregressive Autoregressive
Speed ~70ms for 10s audio ~1s+ for 10s audio
Chinese CER Lower (15-20% improvement) Higher
Punctuation Built-in (ITN) Requires post-processing
VAD Built-in (FSMN-VAD) Separate setup needed

Speed and accuracy numbers are from SenseVoice's published benchmarks.

Features

  • Two input modes: hold-to-talk or toggle (press once to start, again to stop)
  • Auto GPU/CPU selection -- detects VRAM (4GB threshold), falls back to CPU automatically
  • Chinese-first -- optimized for Chinese and Chinese-English mixed input; also supports English, Japanese, Korean, Cantonese
  • Visual feedback -- floating status bar with live volume meter, system tray icon
  • Modular design -- overlay UI is optional and replaceable, backend runs independently
  • Zero native dependencies -- clipboard and key simulation via Win32 API through ctypes, no pywin32 needed

Quick Start

1. Create environment and install PyTorch

# Using micromamba (or conda/venv)
micromamba create -n sensetype python=3.11 -y
micromamba activate sensetype

# GPU (CUDA 12.1)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121

# Or CPU only
pip install torch torchaudio

2. Install dependencies

pip install -r requirements.txt

3. Run

Double-click a launcher script in the project root:

File Description
SenseType.bat With console window, shows logs, good for debugging
SenseType.vbs No console window, double-click and go, best for daily use

Both auto-locate the Python environment and request administrator privileges.

Or run manually from the command line (requires administrator):

python -m sensetype

The model (~400MB) downloads automatically from ModelScope on first run.

Configuration

Edit sensetype/config.py:

Option Default Description
HOTKEY ctrl+alt+z Global hotkey combination
MODE toggle hold = hold-to-talk, toggle = press to start/stop
DEVICE auto auto / cuda:0 / cpu
LANGUAGE auto auto / zh / en / ja / ko / yue
USE_ITN True Auto-punctuation and inverse text normalization
OVERLAY_ENABLED True Show floating status bar (False for headless mode)
SILENCE_THRESHOLD 0.01 Skip audio below this volume level
PASTE_DELAY_MS 80 Delay before restoring clipboard (ms)
AUDIO_SAVE_DIR "" Save recordings to this directory (empty = ~/sensetype_audio)
AUDIO_KEEP_COUNT 10 Number of recent recordings to keep

Architecture

Hotkey Press ──► Recorder ──► Transcriber ──► Paste to Cursor
  (keyboard)    (sounddevice)  (SenseVoice)    (Win32 API)
                    │               │
                    ▼               ▼
               Overlay UI      System Tray
               (tkinter)       (pystray)

The backend pipeline (hotkey -> record -> transcribe -> paste) is fully independent of the UI. Set OVERLAY_ENABLED = False to run without any visual feedback.

sensetype/
├── __main__.py      # Entry point
├── main.py          # Main loop: hotkey listener + orchestration
├── config.py        # All configuration options
├── recorder.py      # Audio recording (sounddevice, 16kHz mono)
├── transcriber.py   # SenseVoice inference via FunASR AutoModel
├── input_paste.py   # Win32 clipboard paste (ctypes, no pywin32)
├── tray.py          # System tray icon (pystray)
└── overlay.py       # Floating status bar with volume visualization (tkinter)

Comparison with Alternatives

Tool Model Local? Relative Speed Chinese Quality
SenseType SenseVoice-Small Yes ~15x faster than Whisper-Large Best (lowest CER)
whisper-writer OpenAI Whisper API No (cloud) Depends on network Good
whisper-typing faster-whisper Yes ~3x faster than Whisper Good
Privox Faster-Whisper + Llama 3 Yes Moderate Good + LLM post-processing
whisper-ptt-windows Whisper (offline) Yes Baseline Good
TalkType Whisper Yes Baseline Good

Note: "Chinese Quality" comparisons are based on SenseVoice's published CER benchmarks against Whisper-Large-v3. TalkType is Linux-only.

Requirements

  • Windows 10/11
  • Python 3.11
  • Microphone
  • NVIDIA GPU with >= 4GB VRAM (recommended) or CPU
  • Administrator privileges (for global hotkey capture)

Roadmap

  • Audio feedback (start/stop recording sounds)
  • Custom text post-processing rules (terminology correction)
  • Auto-start with Windows
  • VRAM stability monitoring for long sessions
  • Overlay UI improvements

License

MIT

Acknowledgements

  • SenseVoice by Alibaba FunAudioLLM -- the ASR model
  • FunASR -- the inference framework