Most datasets are 16khz sampling rate, wav file. A good place to start is Mozilla deepspeech implementation.