Večločljivostno izločanje značilk pri razpoznavanju emocij v govoru : doktorska disertacija : doctoral dissertation

slovenian

doktorska dela

2013

Univerza v Mariboru, Fakulteta za elektrotehniko računalništvo in informatiko

The presented thesis treats the problem of recognizing emotional speech from audio recordings. In order to obtain the optimum processing window width for feature extraction and to achieve the highest level of recognition rates, several different short term processing window widths have been evaluated. In previous studies, the most common process window widths used, were between 20 and 100ms 6, where a shorter window provides better time resolution but poor frequency resolution, while the longer window lifts frequency resolution and worsens time resolution. At this point, we define a new procedure that combines the advantages of narrower and wider windows and takes advantage of dynamically adjusting the time and frequency resolution of individual feature characteristics. The procedure, called ESRA, defines the concept of multi-resolution feature extraction, evaluation and use of selected features to provide for the multi-resolution concept in identifying final classes, where processing acoustic signals, based on the use of voiced speech. Raised levels of recognition performance are achieved using the feature normalization and post feature value smoothing. The added value in the process of optimizing the performance of the recognition algorithm is in defining the algorithm for replacement of final classes, with which increased performance on optimal results have been achieved. To evaluate the impact of the algorithms for optimization of emotional speech recognition, two different areas, called short term and long term areas, are used. On the basis of these two areas the procedures of evaluation and elimination of emotion-dependent features, for the purpose of recognizing emotions in speech, are tested. To confirm the benefits of the algorithm two ways to generate feature subsets and four different classifiers (MLP, RF, KNN, GMM) are used. The audio recordings used are a part of emotional speech database Interface 18, which contains recordings of six basic emotional categories (Ekmans big six) and neutral speech. The highest level of emotion recognition performance with multi-resolution approach was 88.6%, which exceeded the performance of the best single-resolution approach by 3.8%, or by 24,9% speaking in relative terms. There are also comparisons to previous researches recognition rates on above mentioned database.

07.08.2023