(Signal) Loss in Translation?

November 21, 2024

Written by Eric Krapf, General Manager and Program Co-Chair, Enterprise Connect Publisher

Simultaneous, real-time speech-to-speech language translation has become a hot topic. Microsoft made a high-profile announcement in this area during its Ignite conference this week, and most of the other major collaboration platform vendors are also making moves in this space. In theory, it’s a dream as old as civilization: Erase the language barriers that keep diverse people all over the world from communicating. No Jitter has a witty headline hearkening to this ancient yearning with its recent article on the subject, Power of Babel: The Evolution of Real-Time Translation Features.

But that article presents some of the biggest concerns with this emerging capability. For a translation engine, what people say in meetings is essentially data, so the old computer maxim applies: “Garbage in, garbage out.” Like so many AI-driven capabilities, language translation will only be as good as the data it has to work with. No Jitter quotes Snorre Kjesbu of Cisco to this effect: “The main challenge in achieving real-time translations across different languages is ensuring the quality of data being input into the program.”

This means that not only must the AI models performing the translation be of the highest quality, but so must the audio signal coming into the system from each participant’s microphone. The most sophisticated translation engine may hear “Ninety” and be able to translate it into another language in real time, but if the person actually said, “Nineteen,” you’ve got a problem.

Now think about some of the calls you’ve been on. How many included people calling in from the road, on cell phones; or from remote workers with bad Internet connections; or from people at the airport? You’re dealing with a user base whose connections, equipment, and physical environments are likely to be widely divergent.

Collaboration platforms have done a pretty good job reducing background noise in many situations, and AI will certainly be able to continue improving signal quality, which will be important in cleaning up poor-quality audio. But that’s a lot of processing that has to happen really fast. As Sebastian Stüker, director of research science at Zoom explained to No Jitter, low latency is critical to making speech-to-speech translation functional in real-world collaboration settings.

In the days when telephony was still the dominant enterprise communications medium, we used to talk about “enterprise-quality voice.” This meant that a company didn’t want to bet its business on a phone system (and network) unless the quality was good enough to make it unlikely that poor audio quality could lead to costly mistakes or misunderstandings.

Real-time language translation is one of the most exciting applications for AI that I can think of. It really could be a game-changer for almost any enterprise. But for now, it’ll likely be on the list of incredibly promising AI capabilities that need rigorous vetting from the compliance team.

Speaking of AI governance and compliance, we’ve got a session at Enterprise Connect 2025 that addresses just this topic. I hope you’ll check out this session and the rest of our amazing Conference program, and sign up for the event, which takes place March 17 – 20 in Orlando, FL. We’re focused on helping you chart your course through the many challenging decisions around AI and all the other issues that confront your IT organization as you grapple with end user and customer needs around communications and collaboration.

One final note: This newsletter will not publish next week, in observance of the U.S. Thanksgiving holiday. I hope you have a great week.