Research Looking For Flawed Voice Models
Failures in search are not always apparent, but some related to voice can be found by exploring automatic speech recognition. Researchers at the International Computer Science Institute (ICSI) introduced a new project this week focused on exploring ASR to understand the limitations and challenges, and will use the findings to lead new methods for improving the technology. Some improvements in ASR will support voice-related search.
Sponsored by the Intelligence Advanced Research Projects Activity (IARPA) via the Air Force Research Lab (AFRL), the one-year project -- scheduled for completion by March 2013 -- will examine the assumptions behind acoustic modeling. ICSI will share the findings once they are complete.
Roberto Pieraccini, director at ICSI, said speech recognition systems like automated customer services or Apple Siri use the basic voice recognition technology invested in the 1970s. "We're reviewing assumptions made 35 years ago and trying to see how to improve the models," he said.
Acoustic modeling supports ASR by creating statistical representations of each of the distinctive sounds that make up words. This will enable ICSI researchers to discover technical challenges that prevent ASR from being more accurate. The second part of the survey will rely on experts and colleagues who are willing to provide their perceptions on where ASR technology is considered to be most effective, where it fails, and its shortcomings.
As a search engine marketer, would you use a flawed campaign strategy if it leads to success in a roundabout way and allows the brand to build on the campaign?
It's well known that basic algorithms for search recognition have flaws with regard to pronunciation, accents, speed, and noise. "Once you understand what's wrong, the next set of innovations will become more effective," said Nelson Morgan, leader of the speech research activity at ICSI.
It all comes down to simplified assumptions. Morgan said people have used statistical properties for years, knowing they are not really true. "Not because researchers and developers are stupid, but it allows advancements in technology," he said. "In science and engineering, you often make simplified assumptions to move forward. Eventually, you want to refine the models."
The study will reveal the flawed directions in hopes that engineers will fix them. The team has an idea of where to look, but they're not yet sure of the answers.