The latest 20190328 builds of the german models were trained on 412 hours of training material and cover a dictionary of more than 450,000 entries now.

The model was trained on these audio corpora:

voxforge_de         54:42:04
gspv2              126:43:07
forschergeist        1:29:27
zamia_de            18:23:08
m_ailabs_de         36:19:02
cv_de              133:17:49
370 hours
voxforge_de_noisy   17:38:14
voxforge_de_phone   13:40:07
zamia_de_noisy       6:03:23
zamia_de_phone       4:35:15
42 hours

%WER 11.52 [ 26948 / 233903, 1510 ins, 11663 del, 13775 sub ] exp/nnet3_chain/tdnn_250/decode_test/wer_9_0.0
%WER 8.44 [ 19736 / 233903, 983 ins, 10187 del, 8566 sub ] exp/nnet3_chain/tdnn_f/decode_test/wer_9_0.0

You can find download links to all our models and dicts here:

https://github.com/gooofy/zamia-speech#download

comments, suggestions and contributions are very welcome. For more information about the zamia-speech project, please visit http://zamia-speech.org/