How to do scratch training or finetuning for Speaker diarization on my own dataset

[saumyaborwankar]

Please help, I was able to find documentation regarding making my own dataset in nemo format but couldnt find any material to use that dataset to train a model or finetune a pretrained model.

[nithinraok]

Are you planning to train a speaker model or VAD model or both for speaker diarization?

[saumyaborwankar]

Ideally i want to train both.

[nithinraok]

For speaker embedding: Prepare your manifest using these steps: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speaker_recognition/datasets.html#all-other-datasets

Then Prepare the hydra configuration file as detailed here: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speaker_recognition/configs.html#dataset-configuration

You may refer to this section as well on how to use the training script with hydra configuration file.

For VAD: Prepare your manifest using these steps: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/datasets.html#speech-command-freesound-for-vad

Then Prepare the hydra configuration file as detailed here: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/configs.html

Then use this script to train a VAD model.

Inference for speaker diarization:

For speaker diarization inference using trained models refer to these steps.