NMF Based System for Speaker Identification

Automatic identification of speakers from text-independent information is a task required in a broad base of applications. Gaussian Mixture Models are a state-of-the-art solution to the task. We apply this method to a text-independent speech dataset, and present a novel method using Nonnegative Matrix Factorization and sparseness constraints. Results show that our method scores on par with the state of the art, even without optimization, while also providing architecture-related advantages. We provide a comparison of the results for the two methods.