Lawrence Saul, 30
Automated speech recognition has entered daily life via telephones that follow voice commands and PC-based software that transcribes dictation with 97 percent accuracy. That’s good, but for many future applications, it will need to be much better, necessitating advances in the underlying technology. One example: Hidden Markov Models (HMM), the standard technique over the last 20 years.
HMMs, which break speech into timebounded units or states, don’t fully reflect the variability of humans speaking.
In 1997, Lawrence Saul created a more fluid means of expressing duration in speech by plotting it as the length of a curve. The following year, Saul and a colleague at AT&T Labs used this new
method to build a speech recognizer that outperformed the best HMM-based system.
Michael Jordan, an expert on machine learning at the University
of California, Berkeley, says Saul is doing “the most impressive new piece of work in speech recognition in many years. The improved performance he obtains is significant.”