What is hibiki-zero?

Hibiki-Zero is an AI speech translation model built by Kyutai Labs that works in real time. It listens to spoken audio in French, Spanish, Portuguese, or German and produces English speech output as the person is talking, rather than waiting for them to finish and then translating a full recording. The project highlights three qualities it aims for at once: accurate translation, low delay between hearing speech and producing the translated version, and voice transfer, meaning the translated speech attempts to keep something of the original speaker's voice rather than switching to a generic robotic voice. Under the hood, Hibiki-Zero is a 3 billion parameter model, which means it needs a dedicated NVIDIA graphics card to run at a reasonable speed. The project suggests at least 8 gigabytes of video memory will work, with 12 gigabytes being a safer amount for smoother performance. There is no version described for running on a regular CPU or a Mac without an NVIDIA GPU. To use it, you can start a local server with a single command using the uv Python tool, which opens a web interface you can visit in your browser to try real time translation yourself, including an option to expose that interface through a temporary public link. Alternatively, you can run the model directly on existing audio files, translating one or several files in a single batch rather than through the live interface. People without uv installed can instead install the package with pip and run the same server command. The README is intentionally brief and points to a Hugging Face model card, a technical report, an academic paper, and a page of audio samples for anyone who wants to hear examples or read the full research details behind Hibiki-Zero. No license is stated in the README itself.