使用funasr-onnx-online-asr测试paraformer_streaming训练的onnx导出模型出现c++段错误

[QDPeng]

🐛 Bug

• 使用funasr-onnx-online-asr测试paraformer_streaming训练的onnx到处模型出现c++段错误，
• 报错位置：funasr-onnx-online-asr.cpp文件148行，FunASRInferBuffer推理函数
• 使用modelscope上的基础模型能正常运行：speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online-onnx

To Reproduce

• 使用funasr-1.0.19 examples/industrial_data_pretraining/paraformer_streaming目录下的微调脚本训练模型
• 基础模型使用iic/speech_paraformer_asr_nat-zh-cn-16k-common-vocab8404-online，并保持config配置一致
• 模型导出使用export.sh

Code sample

model="outputs"

python -m funasr.bin.export \
++model=${model} \
++type="onnx" \
++quantize=true \
++device="cpu" \
++debug=false

onnx推理

funasr-onnx-online-asr --model-dir outputs --quantize true --audio-fs 16000 --wav-path $wavfile

Environment

• OS : ubuntu22.04
• FunASR Version 1.0.19
• PyTorch Version: 2.2.1
• How you installed funasr : pip
• Python version: 3.10

[lyblsgo]

more logs, more details

[QDPeng]

onnx: 1.15.0, onnxruntime: 1.17.1

more logs, more details

[QDPeng]

more logs, more details

it's config.yaml problem, should write tokens into config.yaml

[lyblsgo]

if the issue persists, you can reopen the issue.