Machine Learning Holds Key to Scaling Up Comcast’s Voice Remote

Comcast’s voice remote for its X1 platform has become an increasingly popular feature, and the operator has kept it stocked with a constant flow of updates that, for example, support voice commands for special events such as the Super Bowl or the Winter Olympics.

The technology that underpins that platform is also branching well beyond the TV. Comcast has already started to integrate the X1 voice remote with Xfinity Home, its home security and automation service.

Related: Comcast Adds AI Smarts to Home Security Cameras

Comcast also recently added another voice wrinkle with a Phone Finder feature for Xfinity Mobile. Those customers can activate it by saying, “Xfinity Mobile, find my phone” into the X1 voice remote, or by uttering their 10-digit phone number into the remote.

That’s expected to be just the tip of the voice-enabled spear. Comcast, which has deployed nearly 20 million voice remotes, envisions it will support voice navigation and other voice-based features across a greater number of “domains” that criss-cross the company’s full lineup of services. But a major challenge comes with creating a system that can do that, at scale, without requiring an army of people to manage and update the system manually.

Progress: It’s All In the Telling
Not only does the system need to know and understand an increasingly broader scope of specific and conversational-style commands, it also needs to grasp the intent of the voice command. Is it a search for a TV program or movie, or is the user telling the smart home system to turn off a light or adjust the thermostat?

Comcast is taking on the scale challenge with a machine learning platform developed in-house that works in tandem with an integrated metadata platform. “We have a really good advantage to work with voice, because of our metadata platform,” Comcast Cable executive director of AI product Jeanine Heck said.

That same metadata system has helped to create a foundation for the voice engine.

Early on, Comcast used a more traditional, pattern-based algorithm that relied on manual tuning. But it later realized that machine learning would be required to train that algorithm to maintain a high level of accuracy while keeping pace with requirements as the scope and complexity of the system continued to expand.

Related: AI, Machine Learning to Change the Customer Experience, Comcast’s Watson Says

Machine learning tied to a language model would also become necessary as the voice platform reaches into more domains while adapting to a broader range of conversation-style commands. “We saw that the machine could learn how to accurately find intent in a better way than it did prior to that,” Heck said, adding that Comcast’s dependence on deep learning to process natural language has only become more pronounced over time. “Our machines can learn how to adapt to those [new domains].”

Added Jonathan Palmatier, Comcast Cable’s vice president of product management, voice control: “You’d need an army of people that are trying to capture every possible way that you can construct a phrase, and [that’s where] it starts to get untenable.”

In addition to expanding voice support to multiple Xfinity services, Comcast has also been working on how voice commands can support customer support functions.