Missouri mathematicians make progress on 'cocktail party problem'

Suppose a law enforcement team was able to bug the home of a suspected terrorist. Then suppose that this terrorist was so popular that every evening dozens of co-conspirators and associated hangers-on came to his house to socialize and, perhaps, discuss their wicked plans. As the total number of voices increased, the neccesity of the bug would increase in inverse proportion to its utility: the more noise, the harder it is to hear any one particular source. Researchers call this the “cocktail party problem,” though terrorists rarely drink cosmopolitans.

Two mathemeticians at the University of Missouri at Columbia claim to have solved the problem, at least theoretically. “Our solution is called ‘signal reconstruction without noisy phase,’” professor Dan Edidin said. “In speech recognition technology, a ‘signal’ could be a recording of 25 people in a room talking at the same time. Our solution shows that we can pull out each voice individually, not just with the words, but with the voice characteristics of each individual. We showed that this ‘cocktail party problem’ is mathematically solvable.” The ability to suss out voice characteristics is a critical advance, allowing law enforcment to make biometric matches from existing datasets in order to identify individual speakers.

Practical applications nevertheless remain elusive. The Missouri mathematicians solved the problem using a neural net rather than a reproducable algorithm. “This isn’t consistent and cannot be duplicated easily,” said professor Peter Casazza. “We need to find a way to design an implementable algorithm that could do this consistently and quickly.”

Both the National Science Foundation and the National Security Agency contributed funding to the research.

-read more in this PhysOrg.comarticle; see also Peter Casazza’s Web page [