Run Time and Accuracy of Machine Learning Algorithms for Speaker Identification

The efficiency and accuracy of machine learning algorithms in embedded systems are paramount for real-time processing tasks such as identifying an intended speaker out of potential sound sources. This paper presents an evaluation of support vector machines (SVMs) and feed-forward neural networks (FFNNs) for the regression problem of speaker identification, with a focus on constant calculation time and the influence of the use of different features. Utilizing a dataset comprising different speaker configurations and combinations of features were examined to assess their influence on accuracy and runtime. Results indicate that SVMs with linear, polynomial, and RBF kernels, as well as FFNNs, demonstrate effectiveness in solving the identification problem by providing real-time capabilities.