Find Audio Needles in Sound Archive Haystacks

Searching for particular words or subjects in thousands of hours of audio or TV broadcast archives or congressional testimony used to be an impossible job. Conventional transcribing and indexing are costly, time-consuming, and subject to error. What if you could get a computer to listen for -- and remember -- the actual sounds of the words?

That's what Fast-Talk Communications, a spin-off of Georgia Tech, says their new technology can do. Just spell a word (or sound it out by breaking it into phonemes, the building blocks of word sounds), and their system can find it buried in a massive audio library.

In a recent article in Technology Review, co-founder Mark Clements gives the example of how to find the word "Sudetenland" in an audio archive about World War II: "[just] sound out what Sudetenland sounds like. Take the name, ‘Sue,’ the city, ‘Dayton,’ and the word, ‘land,’ and string those together, type it in. That gets resolved into the set of phonemes you’re looking for." The system then finds all the spoken references that contain that string of phonemes.

Because it relies on sound, the system cannot distinguish homonyms, such as "root" and "route," or "seize" and "seas." But the system is fast: according to Clements, it processes "on the order of 30 hours of material per second." And it can be trained to adjust for accents and other regional variations in pronunciation.

In addition to news and library applications, Fast-Talk sees potential benefits for call-centers and other types of business users.

Read the ACM abstract or the entire Technology Review article [registration required]