Speaker
Description
This study explores the 'Speech-to-Song Illusion' discovered by Diana Deutsch. Specifically, it questions whether a correlation can be established between subjective human perception of the illusion and objective analyses performed by machine learning tools within the open-source Essentia library and others. To identify and examine the specific region, where machine learning algorithms yield ambiguous classifications (speech - song) for speech recordings, this study utilizes the spectro-temporal degradation techniques developed by Philippe Albouy in 2023. Can the model’s predictions for the same speech recordings, evaluated by human listeners, be correlated with the results of a subjective experiment examining their ability to evoke the speech-to-song illusion?