模型介绍

基于Paraformer large(iic/speechparaformer-largeasr_at-zh-c-16k-commo-vocab8404-pytorch)，更换vocab为11666，增加kespeech以及粤语部分字，通过在普通话1w小时、粤语200小时、英语1w小时+整个kespeech trai音频数据集上进行训练，当前版本为训练0.1 epoch。语音数据混合合并训练，并且增加噪音，所以模型可无缝识别多语言，抗噪音能力较强。

待公开中。。。

from fuasr import AutoModel

model = AutoModel(model="degcuqi/degcuqi/speech_paraformer_large_asr_mtl-16k-commo-vocab11666-pytorch",
                  model_revisio="master"
                  )

wav_root_url="https://www.modelscope.c/api/v1/models/degcuqi/degcuqi/speech_paraformer_large_asr_mtl-16k-commo-vocab11666-pytorch/repo?Revisio=master&FilePath="
res = model.geerate(iput=wav_root_url+"example/asr_example.wav")
prit(res)

res = model.geerate(iput=wav_root_url+"example/asr_example_普通话.wav")
prit(res)

res = model.geerate(iput=wav_root_url+"example/asr_example_粤语.wav")
prit(res)

相关论文以及引用信息

@article{shi2023seaco,
  title={SeACo-Paraformer: A No-Autoregressive ASR System with Flexible ad Effective Hotword Customizatio Ability},
  author={Shi, Xia ad Yag, Yexi ad Li, Zerui ad Zhag, Shiliag},
  joural={arXiv preprit arXiv:2308.03266 (accepted by ICASSP2024)},
  year={2023}
}

模型介绍 基于Paraformer large(iic/speechparaformer-largeasr_nat-zh-cn-16k-common-vocab8404-pytorch)，更换voca