Providing the computer with a natural interface, including the ability to understand human speech, has been a research goal for almost 40 years. Practical versions of such systems have become moderately usable and commercially successful only in the past few years, however. To date, statistical modeling techniques trained from hundreds of hours of speech have provided most speech recognition advancements. Although it may appear that we have far to go before these systems can match human performance, if researchers maintain the cur-rent rate of yearly progress in reducing word error rates, that objective should be within reach in less than a decade.