Unweighted Average Recall Research Articles

Recognizing emotion from Turkish speech using acoustic features

Affective computing, especially from speech, is one of the key steps toward building more natural and effective human-machine interaction. In recent years, several emotional speech corpora in different languages have been collected; however, Turkish is not among the languages that have been investigated in the context of emotion recognition. For this purpose, a new Turkish emotional speech database, which includes 5,100 utterances extracted from 55 Turkish movies, was constructed. Each utterance in the database is labeled with emotion categories (happy, surprised, sad, angry, fearful, neutral, and others) and three-dimensional emotional space (valence, activation, and dominance). We performed classification of four basic emotion classes (neutral, sad, happy, and angry) and estimation of emotion primitives using acoustic features. The importance of acoustic features in estimating the emotion primitive values and in classifying emotions into categories was also investigated. An unweighted average recall of 45.5% was obtained for the classification. For emotion dimension estimation, we obtained promising results for activation and dominance dimensions. For valence, however, the correlation between the averaged ratings of the evaluators and the estimates was low. The cross-corpus training and testing also showed good results for activation and dominance dimensions.

Speaker state classification based on fusion of asymmetric simple partial least squares (SIMPLS) and support vector machines

This paper presents our studies of the effects of acoustic features, speaker normalization methods, and statistical modeling techniques on speaker state classification. We focus on the investigation of the effect of simple partial least squares (SIMPLS) in unbalanced binary classification. Beyond dimension reduction and low computational complexity, SIMPLS classifier (SIMPLSC) shows, especially, higher prediction accuracy to the class with the smaller data number. Therefore, an asymmetric SIMPLS classifier (ASIMPLSC) is proposed to enhance the performance of SIMPLSC to the class with the larger data number. Furthermore, we combine multiple system outputs (ASIMPLS classifier and Support Vector Machines) by score-level fusion to exploit the complementary information in diverse systems. The proposed speaker state classification system is evaluated with several experiments on unbalanced data sets. Within the Interspeech 2011 Speaker State Challenge, we could achieve the best results for the 2-class task of the Sleepiness Sub-Challenge with an unweighted average recall of 71.7%. Further experimental results on the SEMAINE data sets show that the ASIMPLSC achieves an absolute improvement of 6.1%, 6.1%, 24.5%, and 1.3% on the weighted average recall value, over the AVEC 2011 baseline system on the emotional speech binary classification tasks of four dimensions, namely, activation, expectation, power, and valence, respectively.

Application of speaker- and language identification state-of-the-art techniques for emotion recognition

This paper describes our efforts of transferring feature extraction and statistical modeling techniques from the fields of speaker and language identification to the related field of emotion recognition. We give detailed insight to our acoustic and prosodic feature extraction and show how to apply Gaussian Mixture Modeling techniques on top of it. We focus on different flavors of Gaussian Mixture Models (GMMs), including more sophisticated approaches like discriminative training using Maximum-Mutual-Information (MMI) criterion and InterSession Variability (ISV) compensation. Both techniques show superior performance in language and speaker identification. Furthermore, we combine multiple system outputs by score-level fusion to exploit the complementary information in diverse systems. Our proposal is evaluated with several experiments on the FAU Aibo Emotion Corpus containing non-acted spontaneous emotional speech. Within the Interspeech 2009 Emotion Challenge we could achieve the best results for the 5-class task of the Open Performance Sub-Challenge with an unweighted average recall of 41.7%. Further additional experiments on the acted Berlin Database of Emotional Speech show the capability of intersession variability compensation for emotion recognition.