Yan Zhou, Qingkai Fang, Yun Hong, Yang Feng*

Source code and resources for our Findings of ACL 2026 paper: Efficient Training for Cross-lingual Speech Language Models.

This repository contains training and inference code for cross-lingual speech language models (CSLM), together with a Chinese S2S conversation benchmark (BELLE-eval-S2S) used in our experiments.

The current release includes:

• Pre-training and supervised fine-tuning (SFT) scripts
• Inference scripts for general and cross-lingual decoding
• A public Chinese test set for S2S conversation evaluation

Edit paths in cslm/train/pretrain.sh:

• DATA_ROOT / DATA_PATH
• MODEL_DIR
• CACHE_DIR
• OUT_DIR
• distributed environment variables (MASTER_ADDR, MASTER_PORT, etc.) if using multi-node

Then run:

cd cslm/train
./pretrain.sh

Edit paths in cslm/train/sft.sh:

• DATA_ROOT / DATA_PATH
• MODEL_DIR (pretrained checkpoint)
• CACHE_DIR
• OUT_DIR

Then run:

cd cslm/train
./sft.sh

The SFT training code expects JSON/JSONL with fields:

• prompt: user input (string or multi-turn list)
• response: target response (string or multi-turn list)

Minimal single-turn example:

{"prompt": "你好，请介绍一下你自己。", "response": "你好，我是一个跨语言语音语言模型助手。"}

Minimal multi-turn example:

{
  "prompt": ["你好", "请用一句话解释机器学习"],
  "response": ["你好！", "机器学习是让模型从数据中学习规律并用于预测或决策的方法。"]
}

Inference scripts take speech unit sequences as input (one unit sequence per line in a text file).

cd cslm/infer
python decode_general.py \
  --lang zh \
  --unit /path/to/unit_sequences.txt \
  --model-name-or-path /path/to/checkpoint \
  --output-dir /path/to/output

cd cslm/infer
python decode_general_cross.py \
  --lang en \
  --unit /path/to/unit_sequences.txt \
  --model-name-or-path /path/to/checkpoint \
  --output-dir /path/to/output

Outputs are appended to:

• /path/to/output/responses.json

BELLE-eval-S2S is our open Chinese speech-to-speech conversation test set used for evaluation.

• Manifest file: BELLE-eval-S2S/test.tsv
• Audio files: BELLE-eval-S2S/wav

If you use this benchmark, please cite our paper.

This project is built on top of the following open resources:

• Base LLM: meta-llama/Llama-3.1-8B-Instruct
• Speech tokenizer: FunAudioLLM/CosyVoice-300M

Please follow their original licenses and usage policies.

@misc{zhou2026efficienttrainingcrosslingualspeech,
      title={Efficient Training for Cross-lingual Speech Language Models}, 
      author={Yan Zhou and Qingkai Fang and Yun Hong and Yang Feng},
      year={2026},
      eprint={2604.11096},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.11096}, 
}

If you have questions, please contact: zhouyan23z@ict.ac.cn.