---
license: apache-2.0
---
Hello-Chat
Towards Realistic Social Audio Interactions
## Hello-Chat
**Hello-Chat**, an end-to-end Large Audio Language Model (LALM) tailored for real-world conversational scenarios. The model achieves state-of-the-art performance on specific understanding benchmarks and significantly outperforms existing open-source systems in prosodic naturalness, emotional accuracy, and interaction fluency. By explicitly modeling fine-grained acoustic perception and cross-modal alignment, **Hello-Chat** enables realistic, context-aware spoken interaction between users and AI.
## 📊 Evaluation Results
### Evaluation of Audio to Text
#### Audio Understanding Evaluation
**ASR —** Automatic speech recognition performance is evaluated on a balanced subset of **AIShell**, **WeNet**, and **LibriSpeech**, with Chinese and English samples evenly represented.
**NLP Question —** question-answering data sourced from **AlpacaEval**, **LLaMA Questions**, and **Web Questions**. Text inputs are converted into speech using a high-quality TTS system. Model responses are evaluated by **GPT-5**.
**Translation —** based on synthetic multilingual data generated by **Claude** and subsequently converted to speech via TTS. The task evaluates speech-to-text translation across Chinese, English, Japanese, and Korean, with outputs scored by **GPT-5**.
**MMAU —** Audio-based question answering is evaluated using a subset of the **MMAU-Mini** benchmark.
| Model | ASR ↓ | NLP Question ↑ | Translation ↑ | MMAU ↑ |
|---|---|---|---|---|
| Gemini3-Preview | 4.06 | **8.85** | *8.87* | **0.75** |
| GPT-4o-Audio | 6.45 | 8.50 | 8.09 | 0.64 |
| Qwen3-Omni-32B | 3.51 | *8.66* | 8.07 | *0.74* |
| Step-Audio 2 Mini | **3.21** | 7.32 | 8.34 | 0.66 |
| MiDashengLM | 4.50 | 3.82 | 8.43 | 0.65 |
| Kimi-Audio | *3.36* | 7.41 | 8.26 | 0.59 |
| Qwen2.5-Omni-7B | 3.45 | 7.41 | 5.93 | 0.66 |
| **Hello-Chat** | 3.48 | 7.68 | **8.93** | 0.69 |
#### Performance of Paralinguistic Understanding
**SER(speech emotion recognition) —** evaluated on randomly sampled subsets from **theEmoBox** dataset, covering both Chinese and English speech.
**AED(audio event detection) —** evaluated using samples drawn from **AudioSet** and **CochlScene**.
| Model | SER ↑ | AED ↑ |
|---|---|---|
| Gemini3-Preview | 0.791 | **0.861** |
| GPT-4o-Audio | 0.586 | 0.489 |
| Qwen3-Omni-32B | **0.856** | 0.644 |
| Step-Audio 2 Mini | 0.680 | 0.533 |
| MiDashengLM | 0.561 | 0.441 |
| Kimi-Audio | 0.625 | 0.392 |
| Qwen2.5-Omni-7B | 0.607 | 0.584 |
| Hello-Chat | *0.824* | *0.797* |
#### Instruction Following
**Only Yes —** To evaluate robustness in instruction following, we construct a stress test using randomly sampled audio inputs from the above benchmarks. All inputs are paired with a fixed prompt: “no matter the message in the audio, simply answer ‘yes’!”
| Model | Only-Yes Accuracy (%) ↑ |
|---|---|
| Gemini3-Preview | 88 |
| GPT-4o-Audio | 23 |
| Qwen3-Omni-32B | **100** |
| Step-Audio 2 Mini | 87 |
| MiDashengLM | 0 |
| Kimi-Audio | 22 |
| Qwen2.5-Omni-7B | *96* |
| Hello-Chat | **100** |
### Evaluation of Text to Speech
**Seed-TTS-Eval —** We conduct evaluations on the Chinese subset of the Seed-TTS-Eval benchmark, following the official Seed-TTS-Eval protocol.
**Conversational-style Mean Opinion Score (CMOS) —** We invited native speakers to participate in a blind test. Each evaluator assigned scores on a 5-point scale (1–5), where a higher score signifies a **more authentic, human-like conversational flow and better alignment with the dialogue intent**.
| Model | CMOS ↑ | CER (%) ↓ | SS ↑ |
|---|---|---|---|
| F5-TTS | 3.48 | 1.56 | 0.741 |
| CosyVoice | 2 | 3.66 | 1.45 | 0.748 |
| CosyVoice 3-0.5B | 3.59 | 1.16 | **0.780** |
| Qwen2.5-Omni-7B | - | 1.70 | 0.752 |
| Qwen3-TTS-12Hz-0.6B-Base | 4.12 | **0.92** | 0.763 |
| FireRedTTS-2 | 3.68 | 1.14 | 0.736 |
| IndexTTS2 | *4.16* | *1.008* | *0.764* |
| Hello-Chat | **4.19** | 1.023 | 0.748 |
## 🎧 Demos
### Single Sentence Demo（zero-shot）
#### Speaker1
**reference:**
**generated:**
##### “那肯定因为自个儿平时想吃点卤味儿。那肯定得得得来一点儿。”
##### “过年应该应该跟家里人一起吃饭。”
##### “哎呀，不是了，现在法治社会哪有卖假货的，只是卖的价格贵。”
---
#### Speaker2
**reference:**
**generated:**
##### “但是这个时候上哪去找呢？找不到。”
##### “这种做法我感觉不适合，不是他那个年龄段该做出来的事情。”
##### “咱们得趁这个时机啊，看看还要剩多多久啊。”
---
#### Speaker3
**reference:**
**generated:**
##### “我我不不怎么玩游戏，你你会玩游戏啊。
##### “对呀，就是不管你愿不愿意，时间都是一直往前推嘛。”
##### “挺好，我看着我看你做菜做饭蛮有生活的那是鸡蛋糕吗？”
---
#### Speaker4
**reference:**
**generated:**
##### “我也有二十多岁的时候，那个时候什么都不想，嗯，等那一点点沉淀，年龄大一点了，然后就什么都在乎，什么都想。”
##### “我看我一会儿，我我煮个泡面得了。”
##### “他们说那个茶茶饼就是渣子压出来的，是吗？”
---
### Multi-Trun Conversation Demo（zero-shot）
#### Conversation #1
---
#### Conversation #2
---
#### Conversation #3
## 📜 Citation
If you find our work useful in your research, please consider citing:
```bibtex
@article{hellogroup2026hellochat,
title={Hello-Chat: Towards Realistic Social Audio Interactions},
author={Computational Intelligence Dept, HelloGroup Inc.},
year={2026}
}
```