The best sounding limited data models have at least thirty minutes of audio data and have a similar pitch and timbre to a base data set. You can get by with less, but it requires finesse.