State-of-the-art automatic speech recognition (ASR) systems are built using hundreds or even thousands of hours of speech data. Even then, high recognition accuracy is achievable only by carefully constraining the recognition ...