Sparse DNN-based speaker segmentation using side information

Abstract

Sparse deep neural networks (SDNNs) for speaker segmentation are proposed. First, the SDNNs are trained using the side information that is the class label of the input. Then, speaker-specific features are extracted from the super-vector feature of the speech signal by the SDNNs. Lastly, the label of each speech frame is obtained by Kmeans clustering, which is used to segment different speakers of a continuous speech stream. The performance evaluation using the multispeaker speech stream corpus generated from the TIMIT database shows that the proposed speaker segmentation algorithm outperforms the Bayesian information criterion method and the deep auto-encoder networks method.