Training with a dataset other than LibriSpeech

NVIDIA / OpenSeq2Seq

Toolkit for efficient experimentation with Speech Recognition, Text2Speech and NLP

https://nvidia.github.io/OpenSeq2Seq

Apache License 2.0

1.54k stars 371 forks source link

Training with a dataset other than LibriSpeech #446

Closed kourouklides closed 5 years ago

kourouklides commented 5 years ago

Hi, I am trying to train with a new dataset other than LibriSpeech and I have formatted the data-set to be 16KHz wav files and the csv files the same as the ones LibriSpeech format is.

However, train, train_eval and eval modes do not work and they throw an error saying "KeyError: H". Infer mode works and successfully produces a infer-out.txt file. Please let me know what is the problem here.

borisgin commented 5 years ago

Can you attach the complete log file please?

vsl9 commented 5 years ago

Hi, it looks like your ground truth transcriptions contain a capital letter H. We use the following alphabet for LibriSpeech: [" ", "a", "b", ..., "z", "'"]. It can be found in a vocab file. All ground truth transcriptions should contain only characters from the alphabet.

kourouklides commented 5 years ago

Thank you very much. It has been resolved.