Training 8000 kHZ language identification model: Same language during inference #2049
Unanswered
kirillkoncha
asked this question in
Q&A
Replies: 1 comment
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hello!
I am training a model for language identification on 8kHZ audio. During the training, EER falls to 0.10. However, when I am testing model on the same validation set I used during training, the model output is the same language for all the audio files.
I encountered the similar problem during finetuning 16kHZ model. It was solved by shuffling batches. I made sure that batches are shuffled during current training.
It seems to me that the inference could be the problem. I wonder why the documentation of Encoder Classifier states that the audio must be 16000 kHZ?
Beta Was this translation helpful? Give feedback.
All reactions