ASLP-lab/WS-Min-ASR

WenetSpeech-Min ASR Models

ASR models for Minnan transcription, Mandarin transcription, and paired-transcript recognition, described in WenetSpeech-Min. The corpus contains around 10,000 hours of Minnan speech, with paired Minnan and Mandarin transcripts for each utterance.

Available Models

Model	Task	Framework
Qwen3-ASR-WSM-Min	Minnan transcription	ms-swift
Qwen3-ASR-WSM-Mandarin	Mandarin transcription	ms-swift
FireRedASR2-AED-WSM	Paired-transcript recognition	WeNet

Installation

Follow the official installation guides:

• Qwen3-ASR, including the vLLM backend for inference.
• ms-swift for Qwen3-ASR fine-tuning.
• WeNet for FireRed fine-tuning and inference.

Usage

Qwen3-ASR

Select Qwen3-ASR-WSM-Min for Minnan output or Qwen3-ASR-WSM-Mandarin for Mandarin output. Both take Minnan audio as input.

from huggingface_hub import snapshot_download
from qwen_asr import Qwen3ASRModel

if __name__ == "__main__":
    subfolder = "Qwen3-ASR-WSM-Min"  # Or "Qwen3-ASR-WSM-Mandarin"
    model_root = snapshot_download(
        repo_id="ASLP-lab/WS-Min-ASR",
        allow_patterns=[f"{subfolder}/*.safetensors", f"{subfolder}/*.json",
                        f"{subfolder}/*.txt", f"{subfolder}/*.jinja",
                        "examples/G1216S0004.wav"],
    )
    model = Qwen3ASRModel.LLM(
        model=f"{model_root}/{subfolder}",
        gpu_memory_utilization=0.2,
        max_num_seqs=1,
        kv_cache_memory_bytes=1024**3,
        max_model_len=8192,
        max_new_tokens=1024,
    )
    results = model.transcribe(
        audio=f"{model_root}/examples/G1216S0004.wav",  # Or your audio path
        language=None,
    )
    print(results[0].text)

FireRedASR2-AED-WSM

Download the model and run paired-transcript recognition in your WeNet environment:

hf download ASLP-lab/WS-Min-ASR --include "FireRedASR2-AED-WSM/*" \
  --local-dir pretrained_models/WS-Min-ASR
cd pretrained_models/WS-Min-ASR/FireRedASR2-AED-WSM
bash setup_wenet.sh
CUDA_VISIBLE_DEVICES=0 bash infer.sh --model-dir . --task both --beam-size 4

This loads step_149999.pt with its matching configuration and tokenizer/CMVN assets. Use --task minnan or --task mandarin for a single transcript. Add --audio /path/to/audio.wav to use your own audio; otherwise the included example is used. Paired output contains text_minnan and text_mandarin.

Training

• Minnan transcription: Qwen3-ASR-WSM-Min
• Mandarin transcription: Qwen3-ASR-WSM-Mandarin
• Paired-transcript recognition: FireRedASR2-AED-WSM