Speaker Detection in Unlabeled Dataset

It is well known that the main tool of communication between people is speech. Speech separation is the separation of the voices of various speakers or the separation of noise (background noise) from the original audio signal. The implementation of models for speech separation could have a positive effect on many applications. However, reliable speech processing in a real environment often requires automatic speech separation but real data is far from always knowing the number or personality of speakers, and moments in time. The accuracy of the separation of speech for new, previously unknown speakers is worse than the accuracy of the results of a model with supervised learning. In this paper, I try to present an algorithm for finding speakers in an unlabeled dataset, where only audio and information about the number of people participating in it are known.