rumeysakeskin/asr-fine-tuning-for-low-resource-languages

Transfer learning for ASR with subword encoding CTC model (NVIDIA NeMo Citrinet) on low-resource languages

Connect with me:

Languages and Tools:

 

asr-fine-tuning-for-low-resource-languages's Issues

Model Evaluation

Hi,

I trained a Turkish model by applying the steps you described in this repository.
It was very helpful and thanks.

When I evaluate the .nemo model with this code, the output manifest generated is changing Turkish characters to garbage values like "\u00f6 = ö"

Also transcription output is always "pred_test" = "b".

1. Can you suggest a proper way to evaluate my model.
2. And why you think that the prediction is always "b"