Recently, Microsoft brought innovations in audio quality to online meetings on Apple platforms. Now the company boasts that, thanks to “artificial intelligence”, it improves audio on multiple fronts. Last time, it was about eliminating distracting background sounds so that only the speaker’s voice could be heard.
Machine learning model but also suppresses the echo, so you should no longer hear yourself when someone on the other side is listening to the incoming sound into the reducer instead of the headphones. Often, the speaker is closer to the microphone than the speaker, so the incoming sound is louder, making it more difficult to remove.
Especially when both parties are talking at the same time. But when only one speaks, it is more difficult for the other to interrupt. Even that has improved, because now they can both sides listen and talk at the same time. The development team continued to achieve reverberation removal. The captured signal after adjustment sounds as if you were speaking into the microphone for a short distance. You can see the fruits of work in the video below:
Microsoft used roughly to train a model that would handle real-time signal filtering 30,000 hours of speech recordings. It was not data from calls in Teams, but publicly available data, where the team ensured a balance of male and female voice in 74 languages.
To avoid increased complexity, he did not train two models, but one that filters out disturbing sounds and at the same time echoes. This model runs 10% faster than the previous model to suppress noisy sounds. Microsoft eats currently turns on in the Teams client for Windows and macOS and mobile phones will soon be available.
Resources: Microsoft Teams Blog