As the world dealt with the COVID-19 pandemic, many businesses shifted their client calls and meetings online. Global usage of video conferencing apps increased at the height of the outbreak. One of the tools that saw an increase during the pandemic was Google Meets. During the crisis, Google Cloud Chief Thomas Kurian reported a daily usage of 25x higher than it was in January. Google Meet’s daily growth has exceeded 60 percent.
It made free last April as more people worldwide transitioned to remote work arrangements and held meetings online. Google Meet had one flaw: background noise could be heard during video calls. All that is set to change as Google Meet is set to roll out its “denoiser” feature.
Smooth and Less Distracting Calls
Google calls its noise-canceling feature for Google Meet “denoiser.” Although it was already introduced last June 8, the feature will be slowly rolled out, with some users getting it sooner in the month. It will be Web first, followed by Android and iOS devices. The “denoiser” is a feature to filter out background noise during video meetings.
Through a video posted on VentureBeat, Serge Lachapelle, G Suite Director of Product Management, revealed how the feature would work. Although there will still be muffled human voice sounds, there will be no background sound. Lachapelle tested it on a clicking pen, a bag of chips crinkling, and metal clinking inside a glass.
Lachapelle made it clear, however, that not all sounds will be blocked out. Singing, laughter, and breathing can still be heard. He bared that Google will be taking a soft approach first. We don’t want to go beyond and start canceling stuffs that shouldn’t be canceled,” he said. “Sometimes, it’s good for you to hear me taking a deep breath or other natural noises,” Lachapelle continued. The “denoiser” will be an ongoing project with adjustments to be made in order to improve the feature.
Two Years in the Making
According to Lachapelle, the project was based on his own experience in video conference rooms while working in Sweden, and others are in the United States. While meeting with the US team usually around morning in the US and evening in Europe, he could hear clinging sounds and noises of people eating their breakfast or dinner. People are having late meetings and kids screaming. All these distractions triggered the project about a year and a half ago.
But the project goes way back beyond that when Google acquired Limes Audio in 2017. Lachapelle’s team did a lot of work finding the right data, building AI models, and addressing latency.
How The Denoiser Will Work
YouTube videos and engineers played loud musical instruments. Similar to speech recognition, noise-canceling requires training a machine learning model to differentiate human speech and noise. Google also utilized audios from YouTube videos. The algorithm was trained using mixed data that featured noise and clean speech. Google has heavily invested in speech recognition and enhancements, and much of it was reused for the project.
Agreeing on what sounds constitute noise proved to be a challenge for Lachapelle’s team. While most humans agree that unwanted noise during a meeting falls on that category, it’s not going to be easy for an AI model to concur without overdoing it.
Also, the line on what the denoiser will cancel and doesn’t cancel is blurry. It’s not as easy as detecting human voices and blocking everything else. The human voice has such a wide range, Lachapelle said. Screaming is a tough one, as well as dogs barking at certain pitches. For the team, it would still be a work in progress.
User experience considerations
Also, user experience should be considered. Once rolled out, the noise cancellation feature will be set at default. The user will have to switch it off in settings if they want the “noise” to come through. On the Web, there comes three dots at the bottom right, then setting. Under the Audio tab, there is an extra button between the microphone and speakers. You can switch it on or off. This is the noise cancellation feature.
On a typical Google Meet call, the user can mute themselves as well as other users. However, Google chose not to let users cancel out others. The feature will only be available on the sender’s side, where the noise originates. This means that the receiver cannot cancel the noise for what they hear.
Google also decided to put the learning model in the cloud. They wanted to make sure that the feature would be accessible on smartphones and laptops. The team considered that laptops are getting thinner, and loading them too hard with CPU isn’t a good idea.
Although Google also considered using the edge, they opted to go for consistency in experience across devices. By using the cloud, the user experience wouldn’t be the same on every device. There would also be no need to update anything. The feature will be turned on the server side.