Apple at trying to fix a barrier to both language and commerce that currently exists between users and Siri: having the A.I. actually recognize names of local businesses. Its plan? Giving the system access to knowledge of users’ geolocation so it understands the restaurants, boutiques, and hardware stores in specific neighborhoods.
Apple has dubbed this technique “geolocation-based language models,” or Geo-LMs for short. And in a Thursday blog post in the company’s Machine Learning Journal, Siri’s speech recognition team explained they’ll use this to seamlessly bring up information about your local pizza joint instead of responding with the trademark, “Sorry, I didn’t quite get that.”
Siri is a automatic speech recognition (ASR) system that is made up of two units, an acoustic model and a language model. The former captures words and the sounds made when someone speaks, while the former determines how likely a certain series of words are in the language being spoken. In essence, Siri’s at best taking a educated guess at understanding what you’re asking it.
Apple identified two interesting shortcomings. For one, certain words and names used for business aren’t used in standard dialogue, so they seldom appear in the system’s training data and it can’t tell how users will pronounce those words or names.
In an attempt to fix this, Geo-LMs will give the Siri access to what the blog calls “entity names.” This way the voice assistant can better predict when you’re talking about a restaurant and not, say, a park that may share a similar name.
Apple put together its Geo-LM using United States Census Bureau data from 169 areas, which covers 80 percent of the U.S. population. They also built a global Geo-LM for every region not encompassed in those areas. So if a user were to ask Siri to find a local grocery store that wasn’t covered by the census data, it would use the global Geo-LM instead.
While this might fill some gaps in Siri, Apple is still lagging behind the likes of Google that is currently training its next generation voice assistant, Duplex, that is capable of making phone calls in a hyper-realistic human voice. Though it’s important to note that Google is also working towards improving the Google Assistant’s ability to understand various accents, dialects, and languages.
The blog post did not specify when this update will be rolled out. But let’s hope you can ask Siri to order from Denino’s Pizza without having it call your friend Dino by mistake.