Support Super-Vector Machines in Automatic Speech Emotion Recognition

Support Super-Vector Machines in Automatic Speech Emotion Recognition 陳嘉穎 Chia-Ying Chen 國立中山大學資訊工程學系 Department of Computer Science and Information Engineering National Sun Yat-sen University m033040029@student.nsysu.edu.tw 陳嘉平 Chia-Ping Chen 國立中山大學資訊工程學系 Department of Computer Science and Information Engineering National Sun Yat-sen University cpchen@mail.cse.nsysu.edu.tw Abstract In this paper, we use super-vectors in support vector machines for automatic speech emotion recognition. In our implementation, an utterance is converted to a super-vector formed by the mean vectors of a Gaussian mixture model adapted from a universal back- ground model. The proposed method is evaluated on FAU-Aibo database which is well- known to be used in INTERSPEECH 2009 Emotion Challenge. In the case of HMM- based dynamic modeling classifier, we achieve an unweighted average (UA) recall rate of 40.0%, over a baseline of 35.5%, by using the delta features and increasing the number of mixture components. In the case of SVM-based static modeling classifier, we achieve an unweighted average (UA) recall rate of 38.9%, over a baseline of 38.2%, by using the proposed super-vectors. Keywords：Speech Emotion Recognition, GMM, Super-vector, SVM The 2016 Conference on Computational Linguistics and Speech Processing ROCLING 2016, pp. 142-152  The Association for Computational Linguistics and Chinese Language Processing 142 1 Introduction Speech emotion recognition (SER) becomes very popular in recent years [1]. The INTER- SPEECH 2009 Emotion Challenge [2] (henceforth referred to as the Challenge) is a large-scale evaluation plan of SER techniques on FAU-Aibo corpus. In the Challenge, the training set and the test set are defined so fair comparison can be carried out. There are 2 classification mod- els, namely the dynamic modeling of hidden Markov model (HMM) on low-level descriptors (LLDs) and the static modeling of support vector machine (SVM) on supra-segmental feature vectors, which are functional values of sequences of LLDs. In this paper, we focus on the 5-class problem in which a decision among 5 emotional categories has to be made for each test utterance. As published by the organizer of the Chal- lenge, the unweighted average (UA) recall rates of the baseline systems, which use openSMILE toolset for LLD extraction and HTK/Weka toolset for classifiers, is 35.5% for dynamic model- ing HMM and 28.9% for static modeling SVM. Furthermore, as part of the evaluation protocol, when the Synthetic Minority Oversampling TEchnique (SMOTE) [3] is applied to deal with the issue of skewed data, the performance of SVM can be improved to 38.2%. These results will be referred to as the baseline performances. Further progress on FAU-Aibo 5-class problem has been reported over the years after the Challenge. For dynamic modeling, a GMM (equivalent to a one-state HMM) using 13 mel- frequency cepstral coefficients (MFCC) with the first and second derivatives achieves 41.4% UA [4]. A hybrid DBN-HMM system combining deep belief network and hidden Markov model achieves 45.6% UA, which stands as the performance to beat on FAU-Aibo [5]. For static modeling, the anchor model method commonly used in speaker recognition [6] has been transferred to emotion recognition, achieving 43.98% UA with SVM [7]. In this …

Content truncated to 500 words