Speaker Clustering Based on Non-Negative Matrix Factorization Using Gaussian Mixture Model in Complementary Subspace

Speech feature variations are mainly attributed to variations in phonetic and speaker information included in speech data. If these two types of information are separated from each other, more robust speaker clustering can be achieved. Principal component analysis transformation can separate speaker information from phonetic information, under the assumption that a space with large within-speaker variance is a "phonetic subspace" and a space within-speaker variance is a "phonetic sub-space". We propose a speaker clustering method based on non-negative matrix factorization using a Gaussian mixture model trained in the speaker subspace. We carried out comparative experiments of the proposed method with conventional methods based on Bayesian information criterion and Gaussian mixture model in an observation space. The experimental results showed that the proposed method can achieve higher clustering accuracy than conventional methods.