Robust emotional speech classification in the presence of babble noise

Emotional speech recognition (ESR) is a new field of research in the realm of human-computer interactions. Most of the studies in this field are performed in clean environments. Nevertheless, in the real world conditions, there are different noise and disturbance parameters such as car noise, background music, buzz and etc., which can decrease the performance of such recognizing systems. One of the most common noises which can be heard in different places is the babble noise. Because of the similarity of this kind of noise to the desired speech sounds, babble or cross-talk, is highly challenging for different speech-related systems. In this paper, in order to find the most appropriate features for ESR in the presence of babble noise with different signal to noise ratios, 286 features are extracted from speech utterances of two emotional speech datasets in German and Persian. Then the best features are selected among them using different filter and wrapper methods. Finally, different classifiers like Bayes, KNN, GMM, ANN and SVM are used for the selected features in two ways, namely multi-class and binary classifications.