FunASR/egs_modelscope/aishell/paraformer

ModelScope: Paraformer-large Model

Highlight

ModelScope: Paraformer-Large Model

• Fast: Non-autoregressive (NAR) model, the Paraformer can achieve comparable performance to the state-of-the-art AR transformer, with more than 10x speedup.
• Accurate: SOTA in a lot of public ASR tasks, with a very significant relative improvement, capable of industrial implementation.
• Convenient: Quickly and easily download Paraformer-large from Modelscope for finetuning and inference.
  • Support finetuning and inference on AISHELL-1 and AISHELL-2.
  • Support inference on AISHELL-1, AISHELL-2, Wenetspeech, SpeechIO and other audio.

How to finetune and infer using a pretrained ModelScope Paraformer-large Model

Finetune

• Modify finetune training related parameters in conf/train_asr_paraformer_sanm_50e_16d_2048_512_lfr6.yaml
• Setting parameters in paraformer_large_finetune.sh
  • data_aishell: please set the aishell data path
  • tag: exp tag
  • init_model_name: speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch # pre-trained model, download from modelscope during fine-tuning

• Then you can run the pipeline to finetune with our model download from modelscope and infer after finetune:

  sh ./paraformer_large_finetune.sh
  

Inference

Or you can download the model from ModelScope for inference directly.

• Setting parameters in paraformer_large_infer.sh
  • ori_data: please set the aishell raw data path
  • data_dir: data output dictionary
  • exp_dir: the result path
  • model_name: speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch # pre-trained model, download from modelscope
  • test_sets: please set the testsets name

• Then you can run the pipeline to infer with:

  sh ./paraformer_large_infer.sh