Emotion recognition from speech using Gaussian Mixture Model and vector quantization

Abstract

In this paper, there is a demand to evaluate the effectiveness of anchor models applied to the multiclass drawback of Emotion recognition from speech. Within the anchor models system, associate in nursing emotion category is characterized by its line of similarity relative to different emotion categories. Generative models like Gaussian Mixture Models (GMMs) are typically used as front-end systems to get feature vectors want to train complicated back-end systems like Support Vector Machine (SVMs) to enhance the classification performance. There is a tendency to show that within the context of extremely unbalanced knowledge categories, these back-end systems will improve the performance achieved by GMMs as long as associate in nursing acceptable sampling or importance coefficient technique is applied. The experiments conducted on audio sample of speech show that anchor models considerably improves the performance of GMMs by half dozen 2% relative. There is a tendency to be employing a hybrid approach for recognizing emotion from speech that may be a combination of Vector quantization (VQ) and mathematician Mixture Models GMM. A quick review of labor applied within the space of recognition victimization VQ-GMM hybrid approach is mentioned here.