))
Scottish AI: Laughter Detection in Machine Learning
Manage episode 445652460 series 3605861
Innhold levert av Brian Carter. Alt podcastinnhold, inkludert episoder, grafikk og podcastbeskrivelser, lastes opp og leveres direkte av Brian Carter eller deres podcastplattformpartner. Hvis du tror at noen bruker det opphavsrettsbeskyttede verket ditt uten din tillatelse, kan du følge prosessen skissert her https://no.player.fm/legal.
Recognizing laughter in audio is actually a very difficult ML problem, filled with failure. Much like most comedians' jokes. Let's hope some good stuff survives.
This is a review of a student's final year project for a University of Edinburgh computer science course. The project focused on creating a machine learning model to detect laughter in video calls, aiming to improve engagement and reduce muting by automatically unmuting users when laughter is detected. However, the project faced challenges, including poor model performance and the discovery that many non-transcribed regions in the ICSI corpus are not actually silence, but quieter speech by other participants. The student detailed the process of evaluating an existing laughter recognition model, training their own model on the ICSI corpus, investigating the impact of training data on model performance, and examining the practicality of real-time laughter detection. Despite the project's ultimate failure to achieve its original goal, it provided valuable insights, generated a publicly available codebase for future research, and highlighted the importance of analyzing non-transcribed regions in audio data for improved accuracy.
Read Lasse Wolter's paper here: https://project-archive.inf.ed.ac.uk/ug4/20222999/ug4_proj.pdf
71 episoder
Del
