Validation metric trend

[guilinhu]

I first trained an FS-EEND model using train_dia.py on my Libri dataset with 2 speakers, where each speaker's utterance was generated by sampling utterances from the speaker and inserting random silences in between (this will result in a high overlap ratio). However, when I then finetune this model on my Libri dataset with 4 speakers, which is generated in the same way as the 2 speaker dataset, I noticed that the val/speech_falarm and val/speaker_falarm metrics show an rising and then fluctuating trend. The curves can be found below. The grey curve represents model trained on 2 speakers, and the purple curve represents model trained on 4 speakers:

Please ignore the logging frequency since it was changed when training the two models.

Looking forward to your response!

[DiLiangWU]

Hello, this is because during the initial stage of model adaptation, when the model first encounters unseen scenarios, it tends to miss speakers. As a result, the speaker false alarm rate becomes low. As the training progresses, the model becomes more confident in its decisions and starts recognizing the previously missed frames as speakers, which may lead to increases in false alarms or speaker errors. As long as the overall trend of the obj_metric (DER) is decreasing, it is reasonable.