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

transcription error #43

Closed saisrinivas047 closed 6 years ago

saisrinivas047 commented 6 years ago

when i transcribed my audio some how the repeated letters in a word is getting replaced by '1'. example: original: two is a whole lot of our analytics around what customers are looking for based on we mine a lot of data from our website on sales searches on what gets requested in branches so we gotten smarter and smarter about the skus that we add meaning that we are adding things that we are pretty certain have some built in demand

wav2letter: who a whole lot of our own alet around ane part the mo wil1 for be dano when we mind a want a ban ong that night on bel1 dirty on on one any weped it an brachet to wen got omalen about that we had ne1ner1ing thing that would plas certain have dome built into man

vineelpratap commented 6 years ago

This is expected. We use 1 and 2 to denote letter repetitions for example caterpillar -> caterpil1ar.

You can read ASG Criterion section from the paper for more details.

ambigus9 commented 6 years ago

@vineelpratap It's necessary to make manual transcriptions of any audio file to get an Inference of the same audio?

vineelpratap commented 6 years ago

Yes, we currently require it so that we can compute WER. You can keep some dummy values in there for your custom dataset if you don't care about it.

ambigus9 commented 6 years ago

@vineelpratap I don't care about WER, so I just put a word, for example: "HI" on myaudio.trans.txt file, right? Where I must to save this trans.txt file?