Automatic laughter detection in Hungarian spontaneous speech using GMM/ANN hybrid method

Beke, András and Neuberger, Tilda (2013) Automatic laughter detection in Hungarian spontaneous speech using GMM/ANN hybrid method. In: New Perspectives on Speech in Action: Proceedings of the 2nd SJUSK Conference on Contemporary Speech Habits. Copenhagen Studies in Language 43., 2013.03.18-2013.03.20., Koppenhága.

Text neuberger-beke_automatic_laughter_detection_final_vers.pdf Restricted to Repository staff only Download (634kB) | Request a copy

Abstract

The accuracy of a speech recognizer decreases in the case of spontaneous speech because of non-verbal vocalization phenomena such as laughter. The aim of the present research is to develop an accurate and efficient method in order to classify laughter and speech segments at first in Hun- garian spontaneous speech. We used GMM for modeling the data and neu- ral networks for differentiating laughter from other speech events. MFCC algorithm was used to extract the appropriate features, and ROC analysis was carried out to reduce feature dimensions. The corpus for training and testing consists of laughter and speech segments of the largest Hungarian spoken language database called BEA. The high TP rate and AUC value confirmed that the GMM/ANN hybrid system is a particularly good method for solving this problem.

Actions (login required)

Edit Item