Manifold-based incremental community detection method for online speaker identification

摘要

The audio data is always continually increasing in many real applications. Recent works have shown that there was an underlying manifold on which speaker utterances live in the model-parameter space. However, the existing speaker detection methods did not consider both these issues together, which limits their real applications. For this problem, we propose an incremental speaker identification method which makes use of manifold structure. First, we assume there is a network of speeches which already labeled its communities (speakers). For a new coming speech data, we make an initial decision to assign it to the community which includes most of its (k) neighbors within a local (epsilon) neighborhood space, and then update the speech network. After a batch of data, we refine the current structure using label propagation, i.e., we make each node belong to the community on the condition that most of its neighbors belonging to. Using the incremental as well as local determination-based strategy, our method can not only efficiently process the streaming speech data, but also describe the underlying manifold structure of the speech data. Some experiments were conducted on a large Chinese speaker recognition data, and the results confirmed the effectiveness of our new approach.