search

flashlight / wav2letter

Facebook AI Research's Automatic Speech Recognition Toolkit
https://github.com/facebookresearch/wav2letter/wiki
Other
6.39k stars 1.01k forks source link

Libri-clean decoder fails #756

Open AlexandderGorodetski opened 4 years ago

AlexandderGorodetski commented 4 years ago

Question

Hello,

I have just updated my wav2letter docker and I found that my simple decoder based on libri-clean training does not work.

Following is the error message: Invalid dictionary filepath specified

Following is the content of my decode.cfg file:

Decoding config for Mini Librispeech

Replace [...] with appropriate paths

--lexicon=/media_alex/common/projects/wav2letter/db/libri-100/lm/lexicon.txt --lm=/media_alex/common/projects/wav2letter/db/libri-100/lm/3-gram.arpa --am=/media_alex/common/projects/wav2letter/exp/libri-100/librispeech_clean_trainlogs/001_model_lists#dev-clean.lst.bin --test=lists/test-clean.lst --sclite=/media_alex/common/projects/wav2letter/exp/libri-100/decode --lmweight=2.5 --wordscore=1 --beamsize=500 --beamthreshold=25 --silweight=-0.5 --nthread_decoder=4 --smearing=max --show=true

tlikhomanenko commented 4 years ago

@AlexandderGorodetski, you need to fix tokensdir and tokens, I guess it is trying to load the paths you used in the training, you can check them in the log printed on the screen, which values for them are set and if the tokens path exists.

andresy commented 4 years ago

also, a couple of decoding options have been changed (as we improved the decoders) a while ago (for example, silweight is no longer valid, but there is silscore instead).