Voice Activity Detection Through Adversarial Learning

Abstract

Voice activity detection (VAD) plays an important role as a pre-processing block in many speech processing applications like speech coding, speech enhancement, speech recognition systems, etc. The main objective of VAD algorithm is to identify speech and non-speech regions in a given audio signal. However the challenging task for the VAD systems would be classifying speech/non-speech frames in an input audio signal that are corrupted by noise i.e environmental noise. With a view to address such a problem, we propose a new approach to VAD using a deep generative model. These models have the ability to learn the underlying distribution of target data through adversarial learning process. In this work, we explore Speech Enhancement GAN (SEGAN) which is a variant of GAN, to analyze the VAD application. The proposed work is evaluated on a subset of Apollo speech corpus as the dataset contain speech files with multiple challenges such as multiple speakers with different noise types, different Signal-to-Noise Ratio (SNR) levels, channel distortion and non-speech for a long duration. The performance of the system is evaluated using detection cost function(DCF) metric. The proposed work gives a better result when compared to other state-of-the-art methods.