Hibiki

Hibiki is Kyutai's model for simultaneous speech translation: as the user speaks, it generates natural speech in the target language, optionally preserving the speaker's voice, along with a timestamped text translation. It uses a decoder-only multistream architecture derived from Moshi that jointly models source and target speech at a constant 12.5Hz framerate.

The original Hibiki model (2.7B parameters, French-to-English) is released under CC-BY, while the later Hibiki-Zero (3B, trained with reinforcement learning and no word-aligned data) extends to more languages under a non-commercial license. Hibiki stands out as one of the few open models tackling real-time speech-to-speech translation rather than offline transcription.