DS2 small model WER - Githubissues

NVIDIA / OpenSeq2Seq

Toolkit for efficient experimentation with Speech Recognition, Text2Speech and NLP

https://nvidia.github.io/OpenSeq2Seq

Apache License 2.0

1.54k stars 369 forks source link

DS2 small model WER #386

Closed inchpunch closed 5 years ago

inchpunch commented 5 years ago

I ran DS2 small model in evaluation mode without using language model. WER is around 29%. I saw in the default base_params num_epochs is 12. Is this performance expected? Thanks!

borisgin commented 5 years ago

What data set have you used for evaluation?

inchpunch commented 5 years ago

I used librivox-dev-clean, the default dataset in the example_config for small model.

wd929 commented 5 years ago

I got 32% after training. Maybe it does not have any problem due to the small network and dataset compared with the original DS2 and Librispeech.

inchpunch commented 5 years ago

It looks so. BTW that was from default "num_epochs"=12. I ran 28 epochs and then got 25.2%.