Talking Points
THEY ALL SOUND ALIKE
Homonyms--different words that are pronounced the same--can be a pain for speech-recognizers. Some languages are worse than others. English has more than 10,000 possible syllables. Japanese has only 120, which means a vastly larger number of homonyms.
THE MIND REELS
How tough is speech recognition? Consider: a vocabulary of 60,000 words produces 3.6 billion possible two-word sequences.
HI, MOM?
Voice recognition is now widely used for collect calls: Will you accept? Yes or no. AT&T says that saves $100 million a year in saved operator time.
FAMILIAR VOICES
The error rate of speech-recognition systems is decreasing at 30% to 40% a year, thanks to refinements in software algorithms and more affordable computing power.
SHORTCUTS COUNT
Each second shaved off the average connect time by using telephone automated directory-service attendants--the kind that ask you "What city?" and then hand you off to an operator--leads to $1 million in industrywide savings a year.
SAY WHAT?
Understanding syntax only takes computers so far in understanding speech. As part of an exercise in language parsing in 1981, MIT computer scientists crafted one grammatically correct English sentence that had more than two million syntactically correct interpretations.
MYSTERIOUS CHAMBER
One of the reasons it's so hard to synthesize speech is that we still don't fully understand the complex geometry of the vocal tract.
Return to main story
| 
