---
title: Haru-4B
canonical_url: "https://www.modelscope.ai/studios/AI-ModelScope/Haru-4B"
md_url: "https://www.modelscope.ai/studios/AI-ModelScope/Haru-4B.md"
repository: AI-ModelScope/Haru-4B
chinese_name: "Haru-4B · Full-Duplex Voice Chat"
last_updated: 2026-11-16
sdk_type: gradio
sdk_version: 6.29.1
downloads: 0
stars: 1
---

# Haru-4B

> Haru-4B - An interactive demo by AI-ModelScope on ModelScope. Real-time full-duplex voice chat with Haru-4B. Chinese / English interface, custom characters, reference voices, live captions and audio-file testing. Model weights are downloaded from ModelScope…

AI-ModelScope/Haru-4B is an interactive demo (Studio) on ModelScope. built with gradio 6.29.1. also known as 「Haru-4B · Full-Duplex Voice Chat」.

- **Repository**: AI-ModelScope/Haru-4B
- **SDK**: gradio
- **SDK version**: 6.29.1
- **Downloads**: 0
- **Stars**: 1
- **Last updated**: 2026-11-16

Source: https://www.modelscope.ai/studios/AI-ModelScope/Haru-4B

---

# Haru-4B · 实时语音对话 / Full-duplex voice chat

![Haru-4B cover](cover.png)

默认中文，可切换 English。支持边听边说、角色设定、参考音色、实时字幕，以及上传音频测试。

1. 上传或录制一段 3 秒至 3 分钟的单人清晰语音作为音色参考。
2. 设置角色与对话语言，点击「开始对话」。首次使用会加载模型并预热。
3. 点击「实时麦克风」的录音按钮，保持录音即可一边听 Haru 说话、一边接话。建议佩戴耳机。
4. 点击「结束对话」或停止录音，释放当前会话。每次实时对话最长 5 分钟。

也可以切换「上传音频测试」，上传一段想对 Haru 说的话并生成回复。音频测试读取前 30 秒输入，并增加 6 秒静音供模型继续回应。

The interface starts in Chinese and switches to English. Upload or record a clean single-speaker reference (3 seconds–3 minutes), configure a character, start the conversation, then keep the live microphone recording to listen and speak together. Headphones are recommended. The audio-file tab can test the same speech model without microphone access.

Please use a reference voice you have permission to use.

## Deployment

- Model: [higurashi-ai/Haru-4B](https://modelscope.cn/models/higurashi-ai/Haru-4B), downloaded explicitly from `https://modelscope.cn` and cached under `/mnt/workspace`. The international model catalog does not currently include this model. `HARU_MODEL_ENDPOINT` can select a different model source without changing the Studio site.
- Runtime: native **Gradio Blocks**, using the platform Gradio SDK and its queue, microphone streaming, generator audio output, and file endpoints.
- Model audio: continuous 24 kHz PCM, processed in 3200-sample frames (7.5 Hz). Browser microphone chunks are transported every 0.5 seconds; framing carries partial samples into the next chunk.
- Listening and speech output run as separate Gradio events. Microphone input feeds a bounded session queue while an output generator advances the original Haru model and emits WAV chunks plus text.
- The model is loaded lazily during a user-triggered Gradio request. One live conversation uses the model at a time; closing or timing out releases conversation state.
- Inference debug logs report model loading, reference duration, session lifecycle, processed frames, compute time, output sample count, and transcript length. Audio content is not logged.
- `client/` and `haru/haru/server.py` retain upstream source for reference; they do not serve the Studio interface.
- Source: [higurashi-ai/Haru](https://github.com/higurashi-ai/Haru), based on commit `7fe6b7665f8bf035998f401f8cb779fbe53f2751`.
- Model/code: Apache-2.0. Original web client: MIT. See LICENSE and THIRD_PARTY_NOTICES.md.

## Local reproduction

```bash
pip install -r requirements.txt
pip install gradio==6.29.1
python app.py
```

The ModelScope Studio supplies the platform's patched Gradio 6.29.1 SDK automatically.

CPU transport tests can run without model weights or CUDA:

```bash
PYTHONPATH=. pytest -q tests/test_gradio_transport.py
```
