Testing pre-trained model

Deep Speech
#1

Hi, I have downloaded the pre-trained model:
wget https://github.com/mozilla/DeepSpeech/releases/download/v0.5.1/deepspeech-0.5.1-models.tar.gz
tar xvfz deepspeech-0.5.1-models.tar.gz

But when we test it (non-native English speaker), a lot of words get mismatched and they seem not to be in english? Is it supposed to recognize english text correctly of non native speakers?

thanks

#2

It’s not that it is supposed, it’s just that for now, the training dataset is mainly from native speakers, so the model has more difficulties with non native speakers. More contribution to Common Voice in English from non native speakers would help improve that.

#3

@lissyx when downloading the trained model, the size is around 1.8 GB, is this trained model from this dataset which is 30 GB https://voice.mozilla.org/en/datasets

#4

I don’t get your point. v0.5.1 was not trained against Common Voice, because of a mistake, but v0.4.1 was and next v0.6 will be.

1 Like
#5

The size of the model file is dependent upon the parameters chosen, not the size of the dataset, and does not contain raw sound data. So a model trained on 30 GBs of data doesn’t necessarily mean the model will be 30 GBs - it will most likely be a tiny fraction of that.

(In fact, if you look at the size of the files in the download, the model is only around 190 MB and most of the size is taken up by the language model.)