Microsoft: Teams gains noise reduction through AI

[bitheerani90]

Microsoft Teams has officially launched a feature that reduces noise in video calls using artificial intelligence (AI). The feature has been tested since April of this year, shortly after the start of the coronavirus pandemic (Sars-Cov-2). Roughly speaking, the new feature uses AI to identify sounds in a call, differentiating between voice and “noise,” and suppressing the sounds to make the conference clearer.
Typically, video calling apps already uk mobile database built-in noise filtering capabilities, but they mostly work with fixed noises (the sound of a fan spinning, or the air conditioner circulating the air, for example). What Microsoft Teams does is adopt the same intelligent system to filter out sounds that are not constant – think a doorbell ringing, opening a food package, typing on the keyboard or dogs barking.

Teams' noise reduction system works through machine learning, which trained an artificial intelligence model with more than 760 hours of clean speech, comparing it with 180 hours of noise and unexpected sounds. Microsoft used public sound sources to ensure no user's privacy was violated, making the service available in more than 10 languages ​​and balancing the offering between male and female voices.

Teams now reduces over 150 noises
As for the noises tested, Microsoft used publicly available samples of more than 150 types of sounds, including toilets flushing, sleep snoring and other examples. The company noted that an important part of the work involved teaching the system not to recognize emotional sounds, such as laughter and crying, as “noise,” so that they would not be accidentally suppressed.

Finally, the environment where the sounds were applied also became important: in total, there were more than 3,000 real environments and 115,000 synthetic rooms, with varied acoustics to also train the AI ​​model's auditory capacity.

microsoft teams noise reduction artificial intelligence
Microsoft Teams now uses artificial intelligence to eliminate unexpected noise from a video call. Image: Luca Lorenzelli/Shutterstock
“Since we use deep learning, it is important to have a powerful framework for training the model. We use Microsoft Azure to enable us to develop increasingly improved versions of our machine learning model. Another challenge was extracting original speech clean of noise, which had to be done in a way that the human ear would perceive as natural and pleasant,” said an excerpt from the announcement made by Microsoft.

The new feature has already been applied to Microsoft Teams, both in the dedicated app and in the web interface. The update is being made available via the server, so it may take some time for it to appear for all users.