The MIT News reported on April 16th that Pedro Torres-Carrasquillo and his colleagues at Lincoln Laboratory are working on a software that a dialect identification system that compensates for an interpreter's inexperience with multiple variants of a spoken language.
Mostly to use when you are snooping on other people's phones, of course, since an interpreter doing a face-to-face, or even a phone job can always ask the client where they are from.
How are they proposing to do that? Their earlier work on dialect identification focused on building models that mapped the audiowave frequencies of phonemes - the individual sounds of a spoken language. Torres-Carrasquillo, an electrical engineer specializing in speech processing in the laboratory's Information Systems Technology Group, says his group has more recently moved from this phonetic-based approach to lower-level acoustic systems that use the basic spectral similarities of small pieces of spoken utterances. "We are not looking for the types of data linguists deal with - larger units such as phonemes and words," he says. "We're looking at the statistical distributions of basic frequency spectra of small pieces of sounds (...) Our group's idea is that we don't need a model that looks like our data - we need a model that can classify our data," Torres-Carrasquillo explains. "We take very small pieces - snippets of speech - turn them into frequencies, add up all these contributions, and make a model that can tell them apart. We're looking for patterns from just milliseconds of speech."
How is this affected for people with speech impediments, or people like me, whose extensive travel has created a bastardized dialect, and an even more bastardized accent?