GTM-UVigo System for Albayzin 2014 Audio Segmentation Evaluation

TitleGTM-UVigo System for Albayzin 2014 Audio Segmentation Evaluation
Publication TypeConference Paper
Year of Publication2014
AuthorsLópez Otero, P, Docío Fernández, L, García Mateo, C
Conference NameIberspeech 2014: VIII Jornadas en Tecnologías del Habla and IV Iberian SLTech Workshop
Date Published11/2014
AbstractThis paper describes the GTM-UVigo systems for Albayzin 2014 audio segmentation evaluation, which consist on segmentation followed by classification approaches with the same segmentation stage, but different classification approaches. Segmentation is performed by means of a Bayesian Information Criterion (BIC) strategy featuring a false alarm rejection strategy: the process of acoustic change-points is supposed to follow a Poisson process, and a change-point is discarded with a probability that varies in function of the expected number of occurrences in the time interval formed by the previous and candidate change-points. The classifier of the primary system represents the audio segments in a total variability space and then classifies them using logistic regression; contrastive system 1 represents the audio segments by means of Gaussian mean supervectors and classification is performed using a support vector machine; and contrastive system 2 models the different classes with Gaussian mixture models and performs maximum likelihood classification.
Citation Key549