Hey !!!
After training a French model, and generating a output_graph.pbmm. I tested my model but it was not good as all. Is it because I did not use the language model and the tries. If is that the reason, so how can I do that ?
Examples of my model :
WER: 3.000000, CER: 7.000000, loss: 19.907118
- src: “minuit”
- res: “it le huit”
WER: 2.000000, CER: 2.000000, loss: 10.053004
- src: “cherchez”
- res: “che che”
WER: 2.000000, CER: 4.000000, loss: 10.941154
- src: “enfin”
- res: “i can”
WER: 2.000000, CER: 2.000000, loss: 13.227859
- src: “distu”
- res: “dix u”
WER: 2.000000, CER: 5.000000, loss: 13.248935
- src: “laquelle”
- res: “i tell”
WER: 2.000000, CER: 5.000000, loss: 13.811304
- src: “absolument”
- res: “as men”
WER: 2.000000, CER: 5.000000, loss: 14.151725
- src: “lentement”
- res: “late in”
WER: 2.000000, CER: 1.000000, loss: 14.546191
- src: “vingttrois”
- res: “vingt trois”
WER: 2.000000, CER: 5.000000, loss: 15.554608
- src: “bienvenue”
- res: “di menu”
WER: 2.000000, CER: 5.000000, loss: 16.124020
- src: “murmures”
- res: “mir mir”
WER: 2.000000, CER: 5.000000, loss: 16.619673
- src: “qu’importe”
- res: “quam porro”
WER: 2.000000, CER: 3.000000, loss: 18.236885
- src: “dommage”
- res: “de age”
WER: 2.000000, CER: 6.000000, loss: 18.687059
- src: “combien”
- res: “an i”
WER: 2.000000, CER: 4.000000, loss: 18.847189
- src: “extra”
- res: “as a”
I think I have to use my own binaries (for the french model) because the deepspeech take the english binaries by default
So I want to know how can I do that ?
I followed this tutorial : https://github.com/mozilla/DeepSpeech/tree/master/native_client#tensorflow-clone--checkout
but I get error when I execute this command :
bazel build --config=monolithic -c opt --copt=-O3 --copt="-D_GLIBCXX_USE_CXX11_ABI=0" --copt=-fvisibility=hidden //native_client:libdeepspeech.so //native_client:generate_trie
Here is the error : ERROR: Skipping '//native_client:libdeepspeech.so': no such package 'native_client': BUILD file not found on package path
Is it a problem with the bazel version
P.S = I tried bazel 0.19.1 and 0.21.0 and I always get the same error
Thx
