Skip to content

Conversation

@zuazo
Copy link

@zuazo zuazo commented Jan 9, 2023

I have made some more improvements to the previously shared model and #25 :

  1. Trained from scratch with CUDA 11.6 and Tensorflow 1.15.5.
  2. Added Wikipedia corpus to the scorer.
  3. Optimized alpha and beta hyperparameters (134 trials).
  4. Trained on Common Voice 12.
  5. Added EusCrawl corpus to improve the LM.

The new accuracy:

Test Corpus WER CER
Common Voice 12.00% 4.48%

The models can be downloaded from here: https://aholab.ehu.eus/~xzuazo/models/Basque%20STT%20v0.1.7/

@zuazo zuazo mentioned this pull request Jan 9, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant