I wanted to fine tune the existing model. So, the situation is we cannot give same wav files to all three of the data sets i.e. train, dev and test but what if I have the same sentence spoken multiple times by the same person and then having this data in 70%, 20% and 10% distribution, respectively? Because I am looking at a specific area here and not improving the overall model accuracy. Another reason is we only have voice samples of around 10 people.
Please describe more thoroughly your setup, and your data ?