AI BriefingHaru-4B: Full-duplex voice model that listens and speaks in 9 languages every 133ms
higurashi-ai/Haru-4B
About the project
A full-duplex voice conversation model that listens and speaks simultaneously, like a phone call. It accepts user audio every 133ms at a 7.5Hz frame rate and generates its own audio and text within the same time window. The structure ensures that listening and speaking do not wait for each other.
It combines the Qwen3.5-4B text backbone with the VibeVoice audio codec. Without a separate audio transformer or discrete audio tokens, a 4-layer MLP head generates continuous audio in a single step via MeanFlow sampling. It runs in an NVIDIA GPU 12GB environment and provides PyTorch and MLX implementations along with a vLLM-Omni plugin.
It supports 9 languages, including English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean. You can set a character with a text card and specify a voice with a short reference clip to converse. It is a pretrained-based model released under the Apache 2.0 license for research purposes.
higurashi-ai/Haru-4B
The original page has no description.
audio-to-audio
This introduction 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 errors, attribution issues, or removal requests via Contact.