Access the cluster tendency by visual methods for robust speech clustering

Abstract

Identifying the cues for speech segments of speech data is an indispensable task in speaker clustering. The existing techniques perform the task of speech clustering without any prior knowledge of cluster tendency. Many techniques are investigated for finding a prior cluster tendency (CT). During the investigation, the visual access tendency (VAT) is recognized as a reasonable choice to find a cluster tendency. The speech clustering poses three important problems, which are as follows: modelling the speech data, cluster tendency, and effective speech clustering. Modelling is required for defining the shape of the speech segment based on the characteristics of speaker’s voice; hence it is useful for speech recognition. The GMM is a good choice for obtaining the precise model of speech data. Determining the number of speakers (or number of clusters) for the speech is known as cluster tendency. The quality of speech clustering depends on modelling and a prior clustering tendency. The classical algorithms [such as k-means, and minimum spanning tree (MST)-based-clustering] are merged with VAT for determining the effective clustering results along with a prior cluster tendency. We use linear subspace learning for representing the speech segments (or speech utterances) in a projected space of high-dimensional data. Various linear subspace learning techniques are used for improving the speech clustering results. The proposed approaches are hybrid approaches (i.e., k-means-CT, and MST–CT-based clustering), they use expensive steps. For this key reason, we propose another method, direct visualized clustering method, in which we derive the explicit speaker clustering results directly from VAT instead of using either k-means or MST-based clustering. We experimented the proposed methods on TSP speech datasets and done the comparative study for demonstrating the effectiveness of our work.