Exploring the benefits of discretization of acoustic features for speech emotion recognition

Abstract

Abstract We present a contribution to the Open Performance sub-challenge of the INTERSPEECH 2009 Emotion Challenge. Weevaluate the feature extraction and classifier of EmoVoice, ourframework for real-time emotion recognition from voice on thechallenge database and achieve competitive results. Further-more, we explore the benefits of discretizing numeric acousticfeatures and find it beneficial in a multi-class task.Index Terms: speech emotion recognition, discretization offeatures 1. Introduction Emotion recognition from speech has made considerable ad-vances in the last years. The number of research studies ofemotional speech databases has grown, and also first applica-tions and prototypes have been developed [1, 2]. There arelarge EU projects (e.g. Callas 1 and Semaine 2 ) that push real-time emotion recognition. The real-time recognition of emo-tion in speech is also our goal, for which we have developedEmoVoice, our framework for real-time emotion recognitionfrom voice [3], that has already been integrated in a numberof prototypes and showcases. However, real-time processingsometimes requires to accept lower recognition accuracies com-pared to offline research systems. In this contribution to theOpen Performance subchallenge of the INTERSPEECH 2009Emotion Challenge we evaluate our methodology to assess if itis competitive. Since our main focus lies on the acoustic fea-tures, we also explore here whether a discretization of numericacoustic features can make the classification problem easier.Though being promising, discretization has not been investi-gated extensively so far. For instance, Casale and colleagues [4]achieve an improvement by feature discretization on two smalldatabases with acted emotions. Here, we study the effects ofdiscretization on a large database with spontaneous emotionssuch as the Challenge database.The rest of this paper is organized as follows: first, webriefly characterize the challenge database. Next we present ourmethodology to speech emotion recognition, which includesfeature extraction, feature selection and classification. After-wards, we present and discuss our results on the database withtwo pre-processing strategies.