Fun-ASR-MLT-Nano-2512

<div align="center">

⭐ Powered by FunASR — please give us a GitHub Star!

This model is part of the FunASR ecosystem — one industrial-grade open-source toolkit for ASR · VAD · punctuation · speaker diarization · emotion / event · LLM-ASR. A Star really helps the project (and keeps you updated):

🌟 FunASR · 🌟 SenseVoice · 🌟 Fun-ASR · 🌟 FunClip

</div>

Fun-ASR-MLT-Nano-2512

「简体中文」|「English」

Fun-ASR-MLT-Nano-2512 is the 800M-parameter multilingual checkpoint in the Fun-ASR family. It is trained on hundreds of thousands of hours of multilingual speech and supports 31 languages, with emphasis on East and Southeast Asian languages. For the Chinese, English, and Japanese checkpoint trained on tens of millions of hours, use Fun-ASR-Nano-2512.

<div align="center"> <img src="images/funasr-v2.png"> </div> <div align="center"> <h4> <a href="https://funaudiollm.github.io/funasr"> Homepage </a> ｜<a href="#core-features"> Core Features </a> ｜<a href="#performance-evaluation"> Performance Evaluation </a> ｜<a href="#environment-setup"> Environment Setup </a> ｜<a href="#usage-tutorial"> Usage Tutorial </a> </h4>

Model Repository: modelscope, huggingface

</div>

<a name="What's News"></a>

What's New 🔥

• 2025/12: Fun-ASR-MLT-Nano-2512 is the multilingual 800M checkpoint, trained on hundreds of thousands of hours of speech and covering 31 languages.
• 2024/7: FunASR is a fundamental speech recognition toolkit that offers a variety of features, including speech recognition (ASR), Voice Activity Detection (VAD), Punctuation Restoration, Language Models, Speaker Verification, Speaker Diarization and multi-talker ASR.

Core Features 🎯

This repository hosts the MLT checkpoint. Its supported-language scope differs from the standard Fun-ASR-Nano checkpoint.

• 31 languages: Covers the languages listed in the model table above, with focused optimization on East and Southeast Asian languages.
• Compact multilingual model: Uses approximately 800M parameters.
• Language control: Accepts the language names shown in the inference example and supports multilingual recognition.
• Context controls: The FunASR inference API exposes hotwords and inverse text normalization (ITN).

Environment Setup 🐍

python -m pip install -U "funasr>=1.4.1"

# Optional: clone the example repo if you want to run the local demo scripts.
git clone https://github.com/FunAudioLLM/Fun-ASR.git
cd Fun-ASR
python -m pip install -r requirements.txt

<a name="usage-tutorial"></a>

TODO

• Support returning timestamps
• Support speaker diarization
• Support model training

Usage 🛠️

Inference

Using funasr for inference

from funasr import AutoModel


def main():
    model_dir = "FunAudioLLM/Fun-ASR-MLT-Nano-2512"
    model = AutoModel(
        model=model_dir,
        hub="hf",
        trust_remote_code=True,
        remote_code="./model.py",
        device="cuda:0",
    )

    wav_path = f"{model.model_path}/example/zh.mp3"
    res = model.generate(
        input=[wav_path],
        cache={},
        batch_size=1,
        hotwords=["开放时间"],
        # 中文、英文、粤语、日文、韩文、越南语、印尼语、泰语、马来语、菲律宾语、阿拉伯语、
        # 印地语、保加利亚语、克罗地亚语、捷克语、丹麦语、荷兰语、爱沙尼亚语、芬兰语、希腊语、
        # 匈牙利语、爱尔兰语、拉脱维亚语、立陶宛语、马耳他语、波兰语、葡萄牙语、罗马尼亚语、
        # 斯洛伐克语、斯洛文尼亚语、瑞典语 for Fun-ASR-MLT-Nano-2512
        language="中文",
        itn=True, # or False
        llm_kwargs={"do_sample": False},
    )
    text = res[0]["text"]
    print(text)

    model = AutoModel(
        model=model_dir,
        hub="hf",
        trust_remote_code=True,
        vad_model="fsmn-vad",
        vad_kwargs={"max_single_segment_time": 30000},
        remote_code="./model.py",
        device="cuda:0",
    )
    res = model.generate(
        input=[wav_path],
        cache={},
        batch_size=1,
        language="中文",
        llm_kwargs={"do_sample": False},
    )
    text = res[0]["text"]
    print(text)


if __name__ == "__main__":
    main()

Direct Inference

from model import FunASRNano


def main():
    model_dir = "FunAudioLLM/Fun-ASR-MLT-Nano-2512"
    m, kwargs = FunASRNano.from_pretrained(
        model=model_dir,
        hub="hf",
        device="cuda:0",
    )
    m.eval()

    wav_path = f"{kwargs['model_path']}/example/zh.mp3"
    kwargs["llm_kwargs"] = {"do_sample": False}
    res = m.inference(data_in=[wav_path], **kwargs)
    text = res[0][0]["text"]
    print(text)


if __name__ == "__main__":
    main()

<details><summary> Parameter Description (click to expand) </summary>

• model_dir: Model name or local disk model path.
• trust_remote_code: Whether to trust remote code for loading custom model implementations.
• remote_code: Specify the location of specific model code (e.g., model.py in the current directory), supporting both absolute and relative paths.
• device: Specify the device to use, such as "cuda:0" or "cpu".

</details>

Performance 📝

Benchmark scope: The tables below are Fun-ASR family results reproduced from the project report. They do not contain a column identified as Fun-ASR-MLT-Nano-2512, so they must not be interpreted as checkpoint-specific results for this MLT model. MLT per-language results will be added when a reproducible evaluation is published.

The family-level comparison covers open-source benchmarks, Chinese dialect datasets, and industry-specific test sets.

1. Open-Source Dataset Performance (WER %)

Note: Seed-ASR* results are evaluated using the official API on volcengine; GLM-ASR-nano* results are evaluated using the open-source checkpoint.

2. Industry Dataset Performance (WER %)

<div align="center"> <img src="images/compare_en.png" width="800" /> </div>

Citations

@article{an2025fun,
  title={Fun-ASR Technical Report},
  author={An, Keyu and Chen, Yanni and Deng, Chong and Gao, Changfeng and Gao, Zhifu and Gong, Bo and Li, Xiangang and Li, Yabin and Lv, Xiang and Ji, Yunjie and others},
  journal={arXiv preprint arXiv:2509.12508},
  year={2025}
}