AUC optimization for deep learning-based voice activity detection

Abstract Voice activity detection (VAD) based on deep neural networks (DNN) have demonstrated good performance in adverse acoustic environments. Current DNN-based VAD optimizes a surrogate function, e.g., minimum cross-entropy or squared error, at given decision threshold. However, usually works on-the-fly with dynamic threshold, and the receiver operating characteristic (ROC) curve is global evaluation metric for all possible thresholds. In this paper, we propose to maximize area under ROC (MaxAUC) by DNN, which can of terms entire curve. objective AUC maximization nondifferentiable. To overcome difficulty, relax nondifferentiable loss function two differentiable approximation functions—sigmoid hinge loss. study effectiveness proposed MaxAUC-DNN VAD, take either standard feedforward network bidirectional long short-term memory as DNN model state-of-the-art multi-resolution cochleagram Fourier transform feature. We conducted noise-independent training comparison methods. Experimental results show that taking optimization higher than common objectives error cross-entropy. The experimental conclusion consistent across different structures, features, noise scenarios, sets, languages.