I found some examples from a github repo and I tested using the english model and scorer from github and it works almost perfectly. The problem is when I try to use my own model and scorer in another language. The programs recognizes nothing and it performs poorly. I also converted pb to pbmm. What could be the possible issues? The dataset has 12 speakers and almsot 25 hours of speech.
You either need more data or need to use transfer learning.
Thanks!!! I am going to see what I can do
Which language are you trying to train for? Romanian? If so, there is a pretrained model here: https://tepozcatl.omnilingo.cc/ro/
Results are (CER, WER) with and without LM.
You’ll want to use transfer learning from the released English model.
Were did you find the model? Does it also have the dataset? My model is weak because I have only 25 hours of speech which is not enough. How can I do transfer learning from English ? I read the docs and I saw that there is transfer learning in the documentation, but more hints could help me.
I trained the model myself. See e.g. this post 25 hours of speech should be fine for transfer learning. I’m currently retraining Romanian with 100 epochs instead of 50 epochs, which should lead to a slightly better model. But I’m only training with the ~3 hours of training data from Common Voice, so with 25h it should be a lot better.
In terms of how to do transfer learning, it’s described in the documentation, but I use this command.
If you’re interested in more, you can find us on Matrix.
Thanks @ftyers !!! I really appreciate! See you on Matrix