README.md

LiveTrans

English | 中文

Real-time audio translation tool for Windows. Captures system audio via WASAPI loopback, runs speech recognition (ASR), translates through LLM APIs, and displays results in a transparent overlay window.

Perfect for watching foreign-language videos, livestreams, and meetings — no player modifications needed, works with any system audio.

Features

• Real-time translation: System audio → ASR → LLM translation → subtitle overlay, fully automatic
• Multiple ASR engines: faster-whisper, FunASR SenseVoice (optimized for Japanese), FunASR Nano
• Flexible translation backend: Compatible with any OpenAI-format API (DeepSeek, Grok, Qwen, GPT, etc.)
• Low-latency VAD: 32ms audio chunks + Silero VAD with adaptive silence detection
• Transparent overlay: Always-on-top, click-through, draggable — doesn't interfere with your workflow
• CUDA acceleration: GPU-accelerated ASR inference
• Automatic model management: First-launch setup wizard, supports ModelScope / HuggingFace dual sources
• Translation benchmark: Built-in benchmark tool for comparing model performance

Screenshots

English → Chinese (Twitch livestream)

Japanese → Chinese (Japanese livestream)

Requirements

• OS: Windows 10/11
• Python: 3.10+
• GPU (recommended): NVIDIA GPU with CUDA 12.6 (for ASR acceleration)
• Network: Access to a translation API (DeepSeek, OpenAI, etc.)

Installation

1. Clone the repository

git clone https://github.com/TheDeathDragon/LiveTranslate.git
cd LiveTranslate

2. Create a virtual environment

python -m venv .venv
.venv\Scripts\activate

3. Install PyTorch (with CUDA)

Choose the install command based on your CUDA version. See PyTorch official site:

# CUDA 12.6 (recommended)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu126

# CPU only (no NVIDIA GPU)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu

4. Install remaining dependencies

pip install -r requirements.txt
pip install funasr --no-deps

Note: FunASR is installed with --no-deps because its dependency editdistance requires a C++ compiler. The pure-Python alternative editdistance-s is included in requirements.txt as a drop-in replacement.

5. Launch

.venv\Scripts\python.exe main.py

Or double-click start.bat.

First Launch

1. A setup wizard will appear on first launch — choose your model download source (ModelScope for China, HuggingFace for international) and model cache path
2. Silero VAD and SenseVoice ASR models will be downloaded automatically (~1GB)
3. The main UI appears once downloads complete

Configuring the Translation API

Click Settings on the overlay → Translation tab:

Parameter	Description
API Base	API endpoint, e.g. https://api.deepseek.com/v1
API Key	Your API key
Model	Model name, e.g. deepseek-chat
Proxy	none (direct) / system (system proxy) / custom proxy URL

Works with any OpenAI-compatible API, including:

• DeepSeek
• xAI Grok
• Alibaba Qwen
• OpenAI GPT
• Self-hosted Ollama, vLLM, etc.

Usage

1. Play a video or livestream with foreign-language audio
2. Launch LiveTrans — the overlay appears automatically
3. Recognized text and translations are displayed in real time

Overlay Controls

• Pause/Resume: Pause or resume translation
• Clear: Clear current subtitles
• Click-through: Mouse clicks pass through the subtitle window
• Always on top: Keep overlay above all windows
• Auto-scroll: Automatically scroll to the latest subtitle
• Model selector: Switch between configured translation models
• Target language: Change the translation target language

Settings Panel

Open via the Settings button on the overlay or the system tray menu:

• VAD/ASR: ASR engine selection, VAD mode, sensitivity parameters
• Translation: API configuration, system prompt, multi-model management
• Benchmark: Translation speed and quality benchmarks
• Cache: Model cache path management

Architecture

Audio (WASAPI 32ms) → VAD (Silero) → ASR (Whisper/SenseVoice/Nano) → LLM Translation → Overlay

main.py                 Entry point & pipeline orchestration
├── audio_capture.py    WASAPI loopback audio capture
├── vad_processor.py    Silero VAD speech detection
├── asr_engine.py       faster-whisper ASR backend
├── asr_sensevoice.py   FunASR SenseVoice backend
├── asr_funasr_nano.py  FunASR Nano backend
├── translator.py       OpenAI-compatible translation client
├── model_manager.py    Model detection, download & cache management
├── subtitle_overlay.py PyQt6 transparent overlay window
├── control_panel.py    Settings panel UI
├── dialogs.py          Setup wizard & model download dialogs
├── log_window.py       Real-time log viewer
├── benchmark.py        Translation benchmark
└── config.yaml         Default configuration

Known Limitations

• Windows only (depends on WASAPI loopback)
• ASR model first load takes a few seconds (GPU) to tens of seconds (CPU)
• Translation quality depends on the LLM API used
• Recognition degrades in noisy environments or with overlapping speakers

License

MIT License