Anomalous audio event detection: state of the art and trends

Dove e quando

2020-12-10 | Zoom Meeting, Tuesday, December 10th, 2020 - time 3:00 pm

With the development of home assistants, such as Google Home and Amazon Alexa, it is getting amazing to notice how much audio data could be valuable in everyday life. In fact, beyond classical speech recognition and/or synthesis solutions, that have already reached a satisfactory quality, audio event detection offers novel options to improve non-verbal man-machine communication. In fact, audio event detection can outperform visual monitoring for several reasons: (a) Accidents can be signaled in real time, without any need to be monitored by video, (b) Audio data would be available, even in bad luminosity conditions, e.g. at the dark, (c) Alarms could be launched or ignored according to the identification of the type of the event...etc. However, some major problems of anomaly detection in general and anomalous audio event detection in particular are still to be raised. First, the scarcity of anomalous data, in comparison to normal data, makes it difficult to find enough anomalous samples for training. Secondly, anomalous event characterization is rather related to the type of the event, and partly to the acoustic scene, which makes it difficult to find a standard feature set to characterize anomalous audio events. Thirdly, and maybe the most relevant problem, audio events signals in real environments are generally mixed to background noise and/or other events, which may be confusing while detecting the proper event. In this talk, the light will be shed on the major advances in anomalous audio event detection, showing the main techniques and solutions. A special focus will be held on the use of recent machine learning trends to approach this problem, in particular using unsupervised learning either for feature extraction or for anomaly detection BIO: Dr.-Ing. Zied Mnasri, obtained the national engineering diploma in 2003, Msc. degree and Ph.D degree in electrical engineering in 2004 and 2011, from the university of Tunis El Manar in Tunisia. Since Sept. 2011, he has been appointed as assistant professor at the national school of engineering in Tunis, teaching electronics, microcontrollers and digital signal processing. From Jan. 2015 to Dec. 2017, he had been granted the PHC-Utique scholarship to conduct a research project about Arabic parametric text-to-speech synthesis, in cooperation with the LORIA-INRIA, University of Lorraine in France. Since Sept 2018, he has moved to the University of Genoa, Italy, as postdoc researcher. His research interests include audio and speech processing, natural language processing and machine learning. Date: Tuesday, December 10th, 2020 - time 3:00 pm Zoom Meeting: Meeting ID: 917 7863 4392 Passcode: xF91bs
Ultimo aggiornamento 9 Dicembre 2020