DeepSpeech Latest Results with English

DeepSpeech
#1

Hello Team, could you please point me to the link, to refer the latest results for WER for pre-trained English DeepSpeech model?

#2

It’s not yet public. But will be in the next couple of days.

1 Like
#3

you are doing great work guys. keep it up…

#4

Now it’s public.

2 Likes
#5

@kdavis: Please also share the approx training data size in hours used to train the model?

#6

The data sets used are indicated in the release notes linked to in my previous comment.

#7

@kdavis is 0.5.0 only trained on Libreespeech data.
what about 22 gb mozilla open corpus.

#8

No. See the release notes.

#9

hi @kdavis
as per release notes

deepspeech model is trained on American English which achieves an 8.22% word error rate on the [LibriSpeech clean test corpus]

it is mentioned that test set is used as librispeech clean test but not clearly mentioned which train set is used to priovide this checkpoints

please help me clear.

#10

The release notes state

train_files Fisher, LibriSpeech, and Switchboard training corpora.

#11

oh I thought that was for fine tuning. anyway thank you so much…