Speaker: Sara Barahona Quirós. Abstract: In this seminar, we will explore approaches for training acoustic event detection and speaker verification systems employing limited labels. Specifically, for the first task, we will explain the optimization process of a system based on Conformer architecture using different types of labels during training, thus following a semi-supervised learning approach. The results obtained with this system in the international DCASE 2023 evaluation will be presented, as well as an in-depth analysis of the system using a multi-resolution approach to explain the lack of temporal precision of this architecture. Regarding the speaker verification task, an approach using only weak labels during the training of the embedding extractor will be introduced, detailing the steps followed in the process and its implementation within the WeSpeaker framework. Additionally, we will present the preliminary results obtained, which improve the previous implementation in TensorFlow demonstrating the potential of WeSpeaker to achieve more robust training.
