Qwen3-ASR-FT-robohead

Qwen3-ASR-1.7B fine-tuned on the RoboHEARD train partition.

The reported scores are validation results.

Model	Val sessions	cpCER (%)	tcpCER (%)
Qwen3-ASR-1.7B	28	50.95	52.82
Qwen3-ASR-FT-robohead	28	48.56	50.49

Scoring: MeetEval, character-level, tcpCER collar 5 s.

Inference

Install Qwen3-ASR in a PyTorch environment:

pip install qwen-asr==0.0.6 transformers==4.57.6

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "SMIIP-lab/Qwen3-ASR-FT-robohead",
    dtype=torch.bfloat16,
    device_map="cuda:0",
    max_inference_batch_size=1,
    max_new_tokens=256,
)
result = model.transcribe(audio="speech.wav", language="Chinese")
print(result[0].text)

Base model: Qwen Team, Alibaba Cloud. See LICENSE.