This article presents a new method for speaker verification, which is based on the non-negative matrix deconvolution (NMD) of the magnitude spectrogram of an observed utterance. In contrast to typical methods known from the literature, which are based on the assumption that the desired signal dominates (for example GMM-UBM, joint factor analysis, i-vectors), compositional models such as NMD describe a recording as a non-negative combination of latent components. The proposed model represents a spectrogram of a signal as a sum of spectro-temporal patterns that span durations of order about 150 ms, while many state of the art automatic speaker recognition systems model a probability distribution of features extracted from much shorter excerpts of speech signal (about 50 ms). Longer patterns carry information about dynamical aspects of modeled signal, for example information about accent and articulation. We use a parametric dictionary in the NMD and the parameters of the dictionary carry information about the speakers' identity. The experiments performed on the CHiME corpus show that with the proposed approach achieves equal error rate comparable to an i-vector based system.
展开▼
