Higurashi open-sources Haru 4B, a full-duplex speech model trained on 40M hours
alexcovo_eth · x · 2026-11-13
HigurashiAI has open-sourced Haru 4B under Apache 2.0: an end-to-end full-duplex speech-to-speech model (open weights + inference code + technical blog) pretrained on 40M hours of multilingual data including Japanese and English. Unlike cascaded ASR→LLM→TTS systems, Haru 4B processes audio directly and can listen and speak simultaneously, preserving paralinguistic cues like intonation and emotion, and handling backchanneling, interruptions, overlapping speech, and shared silence. Four conversation demos are on the site; the team is also scaling up toward larger models for natural everyday voice AI.
Related event: Higurashi Open-Sources Haru 4B Full-Duplex Speech Model(2 posts)→
More from Multimodal
- Ant Group's Duo v1: a reciprocal world model for real-time full-duplex interaction — inclusionAI · 2026-11-13
- G-STAR from PKU & Meituan LongCat: 20x faster autoregressive image generation at 2.5s per image — meituan-longcat · 2026-11-13
- Creator Uses Seedance 2.5 to Generate a Heartwarming Short Film — SimplyAnnisa · 2026-11-13
- Fable 5.2 composes a symphony mimicking how AI learns, from noise to melody — FinanceYF5 · 2026-11-13
- Ghibli-style playable scene built with Fable 5.2, open sourced — FinanceYF5 · 2026-11-13
- Local pipeline showcase: MiniMax H3 refMod generations from reference image to 4K upscale — edisson75 · 2026-11-13