From August 3, 2023 to November 24, 2023
BUT and Conicet collaborate to advance technologies for multichannel automatic speech recognition (ASR).
Martin Bernardo Meza
The human auditory system has the remarkable ability to extract separate sources from a complex mixture, while this task seems to be difficult for automatic calculation systems, especially when only a monaural recording of mixed-speech is available. During the stay at BUT we will be working on applying speech separation based techniques to enhance the quality of noisy recordings, damaged with specially challenging distortions as late reverberation and non stationary noise sources.
Lautaro Estienne
Indentifying which are the reasons why a Deep Learning model takes a decision could help to develop better and more interpretable models. In collaboration with BUT and CONICET, I am working to better understand the rationales that helps a model to output certain predictions. In addition, it is of our interest to find the relation of these rationales, if any, with the process of calibrating a classification models.
