Hibiki
Kyutai Labs' real-time speech translation — translate as you speak, keep your voice
Kyutai Labs' real-time speech translation system: translates as you speak while preserving the original speaker's voice.
Translate documents, parse PDFs, and extract text with OCR.
Key features
- Streaming French-to-English speech translation
- Adapts flow to translate in real time, chunk by chunk
- Optional voice transfer with controllable classifier-free guidance
- Outputs target speech plus timestamped text translation
Good for
- Developers adding live speech translation to apps
- Mobile users running the small model on smartphones
Stack & tags
PythonPyTorchRustMLXSwiftspeech-translationstreamingsimultaneousvoice-transferopen-weightson-device
Quick start
pip install -U moshi
python -m moshi.run_inference sample_fr_hibiki_crepes.mp3 out_en.wav --hf-repo kyutai/hibiki-1b-pytorch-bf16From the project README on GitHub.
Similar open-source projects
微软 MarkItDown
★ 187,737Microsoft's open-source Python tool — convert PDF/Word/PPT/audio/YouTube to clean Markdown
DeepSeek-OCR
★ 23,926DeepSeek's novel OCR document model — "contextual optical compression"