This repository contains a few modifications to the IO bottleneck encountered with ECAPA-TDNN data loading. After modification, each epoch training time grows up to about 20 minutes.
Added a script ./preprocess/gen_audio.py to augument the training data in advance, the dataloader section has been modified accordingly. This allows the model to reduce the CPU processing time during the dataloader stage.
Before running the trainECAPAModel.py, please run the ./preprocess/gen_audio.py to preprocess your data.Please refer to other more detailed steps ECAPA-TDNN.