Towards Debugging Deep Neural Networks by Generating Speech Utterances