KyutaiHibiki

Hibiki 1B

Hibiki 1B translates French speech into English speech and aligned text as the speaker talks. It aims to preserve the source voice while adapting when enough speech has arrived for a translation.

Research release

Voice fidelity and translation can trade off

The release was trained on recordings up to 120 seconds. Increasing classifier-free guidance can preserve the voice more closely but can reduce translation quality. MLX, Rust and quantized releases are formats of the corresponding model.hibiki-1b-pytorch-bf16: Kyutai model card

Specifications & access

Hibiki 1B specifications
DetailValue and conditions
StatusResearch release
AccessDownloadable
Model identifierskyutai/hibiki-1b-mlx-bf16kyutai/hibiki-1b-pytorch-bf16kyutai/hibiki-1b-rs-bf16kyutai/hibiki-1b-rs-q8kyutai/hibiki-1b-rs-q6k
Reported parameter count1.7 billion parameters; the repository name is a shorter family label.hibiki-1b-pytorch-bf16: Kyutai model card
Audio rate1.1 kilobits per second at 12.5 frames per second.hibiki-1b-pytorch-bf16: Kyutai model card
LicenseCC BY 4.0.hibiki-1b-pytorch-bf16: Kyutai model card

Related models

  • Hibiki Zero 3B

    Kyutai ยท Hibiki

    A speech model that translates French, Spanish, Portuguese and German into English using sentence-aligned training.

    Launch date not confirmedResearch releaseDownloadable