Datasets:
Any stats about improvement in WER going from V1, V2, ..., V5?
#6
by
captain-ballot
- opened
Hello, I'm planning to collect a dataset for another language and would like to know if anyone has any training results to share.
Have you seen diminishing returns to increasing the training dataset from 20 to 30 to 50 to 60 to 110 hours?
Or does the improvement seem linear with respect to training dataset size?