language modeling” (SLM).
The concept, first developed in the 1980s, estimates the probability of how people will group together words, phrases, and sentences according to their natural speech patterns. For Nuance, the company specifically wants to organize the words into “semantic classifications” of meanings. Based on that work, Nuance is developing an “inference engine” that can learn, understand, and interpret voice commands, says Ed Chrumka, senior product manager of connected car services.
Thus, the car will better match the driver’s actual words to their actual intent. In theory, a car will learn a driver’s linguistic habits so eventually there will be no difference between “Call John Smith,” (the original command the car recognizes) and “I wanna call John Smith,” a phrase the driver is more likely to use.
SLM is “a totally different way of doing things,” Richardson says. “It wants to incorporate more natural ways of talking.”
But SLM is easier said than done (pun intended). Human language is complex and broad and entails an almost infinite number of combinations of phrases, words, and sentences. For a computer, recognizing words is the easy part. Figuring out the speaker’s actual meaning is a whole different ball game.
“Perhaps the most frustrating aspect of statistical language modeling is the contrast between our intuition as speakers of natural language and the over-simplistic nature of our most successful models,” Ronald Rosenfeld, a computer science professor at Carnegie Mellon University, wrote in a research paper. “As native speakers, we feel strongly that language has a deep structure. Yet we are not sure how to articulate that structure, let alone encode it, in a probabilistic framework.”
Ford’s job is even tougher. While Google’s voice recognition technology can draw upon the vast resources of cloud-based Internet servers, Ford wants to program its speech software onto a single, self-contained chip installed in the car.
Relying on outside servers will force drivers to spend more money on data plans, Richardson says. Plus it will probably take more time for the car to execute the driver’s command, she says.
However, “you are going to have to have adequate [computing] horsepower in the car to deliver the experience we are all looking for,” Chrumka of Nuance says.
