higurashi

higurashi: Full-duplex speech foundation models

higurashi is a Tokyo startup that researches and builds speech foundation models for natural, human-like conversation with AI. Its first release, Haru 4B, is an end-to-end full-duplex speech-to-speech model that listens and speaks at the same time, published as an open-weight base model with its inference code.

Key Features

• Native full-duplex: Every 133 ms (7.5 Hz) the model takes one frame of the user's audio and produces one frame of its own audio and text, so listening and speaking never wait for each other.
• No cascade: Audio is processed directly instead of chaining ASR, an LLM and TTS, which keeps intonation, emotion and speaking style that are lost in a text step.
• Conversational behavior: Learns backchannels, interruptions, overlapping speech, laughter and turn-taking from real conversations, not from scripted or synthetic dialogues.
• Low latency: higurashi reports about 65 ms end-to-end latency on a MacBook.
• Large-scale training: Pretrained on 40 million hours of real speech across more than 160 languages, with conversations up to 60 minutes long.
• Nine main languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean.
• Character and voice control: A text character card sets the persona, and a reference clip of 3 seconds to 3 minutes sets the voice.
• Efficient design: Built on a Qwen3.5-4B backbone with the VibeVoice continuous speech tokenizer and a small MLP head that generates audio in a single sampling step.
• Ready to run: PyTorch code with a web demo, an MLX version with a native macOS app, and a plugin for serving with vLLM-Omni. A GPU with about 12 GB of memory is enough.

Use Cases

• Speech research: A base model for studying and fine-tuning real-time spoken interaction and turn-taking.
• Voice agent builders: Teams exploring speech-to-speech models as an alternative to cascaded voice pipelines.
• Characters and companions: Expressive conversational characters with a custom persona and voice.
• Multilingual conversation: Natural dialogue in Japanese, English and seven other main languages.

Pricing

Haru 4B weights are free under the Apache License 2.0. It is a pretrained base model, not an instruction-tuned assistant. higurashi has no public price list; for research collaboration, model use, API or product integration, contact contact@higurashi.ai.

Getting Started

• Website: higurashi.ai
• Model weights: Hugging Face
• Inference code: GitHub
• Technical blog: Haru-4B deep dive

higurashi is scaling full-duplex speech models so that talking with AI feels like a natural conversation, and it shares its work openly with researchers and developers.