Speech recognition → translation → voice-cloned text-to-speech → dubbing → a finished video with subtitles. No cloud services, no per-minute fees.

🎙️ ASR  →  🌐 LLM Tranlation  →  🗣️ TTS  →  🎚️ dubbing  →  🎬 video + SRT

• 🌍 Any language pair — auto-detects the source, dubs into your target
• 🗣️ Voice cloning keeps the original speaker's voice (OmniVoice; XTTS / F5-TTS / Qwen3-TTS / ElevenLabs also supported)
• 🎚️ Lip-/timing-aware dubbing with speedup and stretch_whole algorithms
• 🖥️ One container, one GPU — runs entirely local, browser UI included
• 📝 Outputs the dubbed video plus source & translated SRT subtitles

git clone https://github.com/langswap-app/langswap
cd langswap

docker build -t langswap .
docker run --rm --gpus all -p 7860:7860 \
  -e HF_TOKEN=$HF_TOKEN \
  -v "$PWD/models_weights:/models" \
  -v "$PWD/data:/app/data" \
  langswap

Open http://localhost:7860, drop in a video, pick a target language. Done.

Needs: an NVIDIA GPU + the NVIDIA Container Toolkit. First run: models auto-download into ./models_weights on first use (set HF_TOKEN for gated models like Gemma). Pre-place weights there to skip the download.

uv venv --python 3.12 && source .venv/bin/activate

uv pip install -e ".[gpu]"     # full local-model stack (needs an NVIDIA GPU)
# or, on a Mac / no GPU — relies on hosted APIs, far fewer deps:
uv pip install -e ".[api]"

python gradio_demo.py                          # browser UI  → http://localhost:7860
python main.py local in.mp4 english russian    # CLI, stage-by-stage

The full GPU install runs everything in one process on transformers 5.x — ASR included. See docs/advanced.md for the exact GPU install (torch cu130 + qwen-asr/qwen-tts).

docs/advanced.md — model list & overrides, exact pinned install, every env var, Docker build notes, and troubleshooting.

Code is AGPL-3.0-or-later (see LICENSE) — run a modified version as a network service and you must publish your source. A commercial license is available: ilya@langswap.app.

Model weights are licensed separately and some (Gemma, Qwen, OmniVoice) restrict commercial use — you're responsible for complying with each model's terms. The AGPL grant on this code does not cover the weights. Contributions are governed by CONTRIBUTING.md.