higurashi · Full-duplex speech model

Haru 4B

Haru 4B is an end-to-end full-duplex speech-to-speech dialogue model from higurashi that listens and speaks at the same time. It processes audio directly rather than through a cascaded ASR, LLM and TTS pipeline, and is meant to handle backchanneling, interruptions and overlapping speech. It aligns user speech, model speech and text streams at 7.5 Hz (about every 133 ms) and uses the continuous speech tokenizer from VibeVoice. The pretrained base model (not an instruction-tuned assistant) is released as open weights on Hugging Face, with inference code on GitHub; the page gives no licence details or benchmark figures.

Open the source ↗

Similar Audio & Speech releases

  • MOSS-Live

    模思智能 (MOSS) · Full-duplex voice agent · Week 46, 2026

  • Dessa

    Block (Dessa research lab) · Speech-only conversation model · Week 45, 2026

  • Gemini 3.8 Live

    Google · Live Voice Agent · Week 39, 2026

  • VoiceMem

    Tsinghua University · Voice Memory · Week 35, 2026

  • ACE Music v1

    ACE Music · Vocal music generation · Week 47, 2026

  • ElevenAI

    ElevenLabs · Meeting Dictation App · Week 47, 2026

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.