Model selection of combined neural nets for speech recognition