AI BriefingHaru-4B: Low-Latency Multilingual Full-Duplex Spoken Dialogue Model Released
·2026.11.15 22:58
Key point
The 4B parameter model processes audio and text in 133 ms frames using a single sampling step for continuous dialogue.
Details
- Haru-4B is a pretrained base model released for research, designed as a low-latency, multilingual full-duplex continuous spoken dialogue foundation model.
- The model operates natively at 7.5 Hz, processing one frame of user audio and producing one frame of model audio and text every 133 ms.
- Audio is represented as a 64-dimensional latent per frame and generated in a single sampling step via a small MLP head with MeanFlow, eliminating the need for separate audio transformers or discrete audio tokens.
- Supports nine languages including English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean, with character and voice control via text cards and reference clips.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Block releases Dessa, a full-duplex speech model trained exclusively on human audio
Hugging Face Hack Reveals Frontier AI Alignment Failure, Not Cybersecurity Gap