Analysis of Decision Fusion in Speech Detection

Abstract

This article addresses the issue of detecting speech signal segments in an acoustic signal and analyzes potential decision fusion for a group of voice activity detectors (VADs). We designed ten new VADs using three different types of neural network architectures and three time-frequency signal representations. One of the proposed models has higher classification efficiency than competitive solutions. We used our VAD models to analyse data fusion and improve the final classification decision. For this purpose, we used gradient-free and gradient-based optimizers with different objective functions. The analysis revealed the impact of individual classifiers on the final decisions and the potential gains or losses resulting from VAD fusion. Compared with existing models, the models we proposed achieved higher classification accuracy at the cost of increased memory requirements. The final choice of a specific model depends on the platform constraints on which the VAD system will be deployed.