modelscope-funasr speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404模型的热词功能同一批wav数据用pipeline推理有效，用ContextualParaformer推理就无效是什么原因。win11+funasr0.8.4+modelscope1.9.5

1 热词有效

hotword_dict = {'hotword': '访谈 开始访谈'}
inference_pipeline = pipeline(
    task=Tasks.auto_speech_recognition,
    model="damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404",
    param_dict=hotword_dict)
funasr_result = inference_pipeline(audio_in=wav_path)

# 2 热词无效
model_dir = "damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404"
model = ContextualParaformer(model_dir, batch_size=1)
funasr_result = model(wav_path, "访谈 开始访谈")

补充细节：
win11+torch1.13.0+funasr0.8.4+modelscope1.9.5，同一批wav数据下：1、https://alibaba-damo-academy.github.io/FunASR/en/egs_modelscope/asr/TEMPLATE/README.html按这个链接的Paraformer-contextual Model示例热词效果较好，是pipeline的推理方式。2、https://github.com/alibaba-damo-academy/FunASR/blob/main/runtime/python/onnxruntime/demo_contextual_paraformer.py按这个链接的热词效果几乎没有，是用ContextualParaformer推理。