---
base_model:
- Qwen/Qwen2.5-Omni-7B
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: audio-text-to-text
---

# EmotionThinker: Prosody-Aware Reinforcement Learning for Explainable Speech Emotion Reasoning

This repository contains the model presented in the paper [EmotionThinker: Prosody-Aware Reinforcement Learning for Explainable Speech Emotion Reasoning](https://huggingface.co/papers/2601.15668).

[![ICLR 2026 Oral](https://img.shields.io/badge/ICLR%202026-Oral-gold)](https://arxiv.org/pdf/2601.15668) [![Project](https://img.shields.io/badge/Project-Page-green)](https://github.com/dingdongwang/EmotionThinker)

<p align="center">
  <img src="intro.png" width="800"/>
</p>

## Introduction
EmotionThinker is the first RL–enhanced SpeechLLM framework for interpretable speech emotion reasoning. For details, please refer to the [paper](https://huggingface.co/papers/2601.15668). 

Unlike conventional speech emotion recognition (SER) systems that treat emotion as a flat classification problem, EmotionThinker reframes SER as a deep reasoning problem, enabling models to jointly produce accurate emotion labels and structured, human-aligned explanations.

EmotionThinker offers the following advantages:

- Higher emotion recognition accuracy compared to existing SpeechLLMs;
- Deep reasoning ability to integrate emotion-related cues for justification;
- Fine-grained audio caption covering speaker traits, prosodic cues and semantic information.


## Quickstart

```python
import torch
from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor
from qwen_omni_utils import process_mm_info

processor = Qwen2_5OmniProcessor.from_pretrained('ddwang2000/EmotionThinker')

model = Qwen2_5OmniForConditionalGeneration.from_pretrained('ddwang2000/EmotionThinker',torch_dtype="auto", device_map="auto")

print("✅ Model loaded successfully")

audio_path="angry.wav" #your audio path
prompt="<audio>What is the emotion expressed in this audio clip? Please choose one from the following options: neutral, happy, sad, angry, contempt or disgust, confused, whisper, surprise, fear."

messages = [
    {"role": "system", "content": [
        {"type": "text", "text": "A conversation between User and Assistant. The user asks a question, and the Assistant solves it. The assistant first thinks about the reasoning process in the mind and then provides the user with the answer. The reasoning process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think><answer> answer here </answer>."} 
    ], },
    {"role": "user", "content": [
        {"type": "audio", "audio": audio_path},
        {"type": "text", "text": prompt},
    ]
     },
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
audios, images, videos = process_mm_info(messages, use_audio_in_video=False)
inputs = processor(text=text, audio=audios, images=images, videos=videos, return_tensors="pt", padding=True, use_audio_in_video=False)
inputs = inputs.to(model.device).to(model.dtype)

with torch.no_grad():
    text_ids = model.generate(
        **inputs,
        return_audio=False,
        max_new_tokens=2048
    )[:, inputs.input_ids.size(1):]


text = processor.batch_decode(text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
print(text)
```

## Citation
If you find this model useful in your research, please kindly cite:
```bibtex
@inproceedings{wang2026emotionthinker,
  title={EmotionThinker: Prosody-Aware Reinforcement Learning for Explainable Speech Emotion Reasoning},
  author={Wang, Dingdong and Liu, Shujie and Zhang, Tianhua and Chen, Youjun and Li, Jinyu and Meng, Helen},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}
```