Create pre-trained model on Librispeech
SeanNaren opened this issue · comments
There has been a lot of attempts to get the model trained on Librispeech 1k hours of training data so providing a trained model I think would be very beneficial (as well as the steps to replicate).
- Prepare dataset on file system
- Create LMDB and make modifications to support this
- Train till convergence
- Add model to repo and document
Stretch goal:
- Allow loading of previous model and training from these weights
+1
I've added a checklist to the main issue which I'll update as I complete tasks!
I'm training the model now. I've modified params and will explain how I trained this model from the ground up. It doesn't strictly follow the DeepSpeech architecture mainly out of performance reasons and aiding convergence.
Hi @SeanNaren, are you planning to upload the model? The community thanks you!
@iassael Sorry for the disappearance... will upload the models tomorrow and give the WER/CERs currently for them. Still plan on improving the smaller version but will get the release out!
@SeanNaren that's amazing. If you need any help with hosting let me know!
Here we are guys :) Huuuge thank you to @maciejkorzepa for giving me the LibriSpeech model!
I think we could squeeze a little bit more out of some of the models in terms of accuracies, but they are a good baseline to start from!
@iassael Thanks for the offer of hosting, Seems like github handled them alright
@SeanNaren do you have information on how you were able to make this converge?