Obtain per-word confidence score

Hello everyone!

Sorry, if i’ve duplicated the question (hope no), but I haven’t found the answer myself.
I want to obtain per-word confidence scores during inference. All what I get so far is json with total sentence confidence using –json key for deepspeech prebuilt binary (v. 0.6.0.alpha.5-50).
What I want instead - confidence score for each word. Is it possible anyhow?
Anyway, what that score exactly means and how can it be used? Since it is very differ from sentence to sentence (I believe depending on number of words), I don’t understand it’s value.
The second thing is, word-timing is always starts at 0, though has different duration. What does it stand for?
Thank you a lot!

(deepspeech2) root@neuro:/opt/deepspeech2/new_DS/DeepSpeech/native_client# ./deepspeech --model $p/models/Deepspeech/output_graph.pb --alphabet /ssd/ru_speech_ds/lang/ru_alphabet.txt --lm $p/lang/lm.binary --trie $p/lang/trie --json --audio /home/some_user/t1.wav
TensorFlow: v1.14.0-14-g1aad02a
DeepSpeech: v0.6.0-alpha.5-50-g90c2acd
Warning: reading entire model file into memory. Transform model file into an mmapped graph to reduce heap usage.
2019-09-12 16:02:13.938085: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
{“metadata”:{“confidence”:451.896},“words”:[{“word”:“а”,“time”:0,“duration”:0.28},{“word”:“д”,“time”:0,“duration”:0.7},{“word”:“пеленой”,“time”:0,“duration”:1.46},{“word”:“другом”,“time”:0,“duration”:1.98},{“word”:“по”,“time”:0,“duration”:2.16},{“word”:“простоте”,“time”:0,“duration”:2.76},{“word”:“так”,“time”:0,“duration”:3.14},{“word”:“как”,“time”:0,“duration”:3.42},{“word”:“выбрать”,“time”:0,“duration”:3.94},{“word”:“неуплата”,“time”:0,“duration”:4.86},{“word”:“поленовы”,“time”:0,“duration”:5.52},{“word”:“ультраправая”,“time”:0,“duration”:7.1},{“word”:“именно”,“time”:0,“duration”:7.58},{“word”:“полипропилена”,“time”:0,“duration”:8.36},{“word”:“утих”,“time”:0,“duration”:8.98},{“word”:“а”,“time”:0,“duration”:9.02},{“word”:“участком”,“time”:0,“duration”:9.7},{“word”:“должен”,“time”:0,“duration”:10.42},{“word”:“или”,“time”:0,“duration”:10.66},{“word”:“отопления”,“time”:0,“duration”:11.18},{“word”:“а”,“time”:0,“duration”:11.34},{“word”:“уж”,“time”:0,“duration”:11.8},{“word”:“подходит”,“time”:0,“duration”:12.32}]}

Currently, the API only gives confidence over the whole transcription. Getting something different would require more and changing the API.