For a long time, speech technology has been a dangling carrot that keeps getting yanked away. With infinitely iterative syntax, omitted-but-implied subjects and actions, and meaning that radically changes based on context, human language is both confoundingly complex and so simple that “even a child can do it.” As a species, we’ve used the tool of language for millennia, despite a shallow understanding of how it actually works. The gut feeling that language is easy and natural is misleading, since it turns out speech technology is one of the tech world’s most grueling challenges.
Yet as the internet becomes more and more entwined with our daily lives, it gets less practical to continue using keyboards as an intermediary. Engineering teams are painstakingly trimming away milliseconds of processing time; relying on such slow input as type is incongruous.
Here’s the amazing news. After decades of speculation that dependable speech technology is “just around the corner,” we’re finally staring that reality in the face. Google recently revealed that the word error rates in their speech recognition technology decreased by 313% in just two years. Other deep-pocketed tech giants are reporting similar bursts of improved accuracy, such as China’s Baidu:
Word error rates in clean, noisy, and combined environments. Credit: Baidu
Users are taking note. In a study, Thrive Analytics found technology improvements to be the leading driver of increased virtual assistant usage (which nearly doubled over a recent 12-month period). The popularization of speech recognition technology is self-perpetuating, since technological improvements are fueled by accumulated speech data. This train is accelerating fast.
Interested in how meaning is derived from speech recognition data? Check out MindMeld’s platform for voice-driven content discovery.
