Views
No views yet
| Attribute | Value |
|---|---|
| Architecture | Qwen3ASRForConditionalGeneration |
| Base Model | Qwen3-ASR-0.6B |
| Parameters | ~0.6B |
| Dtype | bfloat16 |
| Audio Encoder | Whisper-like (18 layers, d_model=896) |
| Text Decoder | Qwen3 (28 layers, hidden=1024) |
pip install -U qwen-asr1import torch
2from qwen_asr import Qwen3ASRModel
3
4model = Qwen3ASRModel.from_pretrained(
5 "TaurenMountain/FormalASR-0.6B",
6 dtype=torch.bfloat16,
7 device_map="cuda:0",
8 max_new_tokens=512,
9)
10
11results = model.transcribe(
12 audio="your_audio.wav",
13 language="Chinese",
14)
15
16print(results[0].text)1import torch
2from modelscope import snapshot_download
3from qwen_asr import Qwen3ASRModel
4
5# 下载模型到本地（首次运行自动下载）
6model_dir = snapshot_download("TaurenMountain/FormalASR-0.6B")
7
8model = Qwen3ASRModel.from_pretrained(
9 model_dir,
10 dtype=torch.bfloat16,
11 device_map="cuda:0",
12 max_new_tokens=512,
13)
14
15results = model.transcribe(
16 audio="your_audio.wav",
17 language="Chinese",
18)
19
20print(results[0].text)1@misc{ning2026formalasrendtoendspokenchinese,
2 title={FormalASR: End-to-End Spoken Chinese to Formal Text},
3 author={Wanyi Ning and Yinshang Guo and Haitao Qian and Jiyuan Cheng and Wei Zhou and Weiyuan Feng and Yufei Zhang},
4 year={2026},
5 eprint={2605.19266},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2605.19266},
9}