Application for Real-time Personalized Speaker Extraction

This short paper demonstrates an audio processing desktop application that allows isolating in real-time the voice of a specific speaker from the possibly noisy audio input after a short enrollment phase. The machine learning model embedded in this application suppresses all other sounds than the target voice from the incoming audio stream, including disturbing distractor voices. In the context of a growing need for video-collaboration solutions, personalized speech enhancement enables the use of such technologies in more challenging acoustic environments, i.e., in the presence of near distractor speech. In this situation, classical speech enhancement systems typically fail as they do not filter out any speech, hence the need for personalized methods. The presented application is an all-in-one solution for personalized speech enhancement: it allows the user to enroll and then to apply the effect seamlessly for one-to-one or one-to-many online meetings.