---
license: gpl-3.0
dataset_info:
  features:
  - name: split_name
    dtype: string
  - name: index
    dtype: int64
  - name: round
    dtype: int64
  - name: question
    dtype: string
  - name: question_audio
    struct:
    - name: array
      sequence: float32
    - name: path
      dtype: string
    - name: sampling_rate
      dtype: int64
  - name: answer
    dtype: string
  - name: answer_cosyvoice_speech_token
    sequence: int64
  - name: answer_snac
    dtype: string
  splits:
  - name: train
    num_bytes: 800059817200
    num_examples: 1400398
  download_size: 792877562556
  dataset_size: 800059817200
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
task_categories:
- question-answering
language:
- zh
size_categories:
- 1M<n<10M
---

# Belle_1.4M

*This dataset is prepared for the reproduction of [SLAM-Omni](https://arxiv.org/abs/2412.15649).*

This is a **multi-round Chinese spoken dialogue** training dataset. For code and usage examples, please refer to the related GitHub repository: [X-LANCE/SLAM-LLM (examples/s2s)](https://github.com/X-LANCE/SLAM-LLM/tree/main/examples/s2s)


## 🔧 Modifications

1. **Data Filtering**: We removed samples with excessively long data.

2. **Speech Response Tokens**: We used [CosyVoice](https://github.com/FunAudioLLM/CosyVoice) to synthesize corresponding semantic speech tokens for the speech response. These tokens, represented as `answer_cosyvoice_speech_token`, are included as model training targets.

3. **User Instruction Speech**: Synthesized speech for user instructions using CosyVoice, with timbres randomly selected from 1,010 Chinese prompts in the [seed-tts-eval](https://github.com/BytedanceSpeech/seed-tts-eval) subset to ensure diversity.


## 🙏 Acknowledgment

The original dataset was sourced from [Belle_train_3.5M_CN](https://huggingface.co/datasets/BelleGroup/train_3.5M_CN). We thank the Belle Group for their open-source contribution.


## 📄 Citation

If you find our work helpful, please consider citing:

```bibtex
@article{chen2024slam,
  title={SLAM-Omni: Timbre-Controllable Voice Interaction System with Single-Stage Training},
  author={Chen, Wenxi and Ma, Ziyang and Yan, Ruiqi and Liang, Yuzhe and Li, Xiquan and Xu, Ruiyang and Niu, Zhikang and Zhu, Yanqiao and Yang, Yifan and Liu, Zhanxun and others},
  journal={arXiv preprint arXiv:2412.15649},
  year={2024}
}
```