Semi-supervised model selection with applications to speech recognition