Speech Activity Detection Under Adverse Conditions Using Neural Networks and Speaker Diarization

Abstract

In this study, a robust speech activity detection (SAD) system consisting of a neural network classifier and a speaker diarization module that works under noisy conditions is designed. The performance of baseline SAD systems degrades significantly under noisy conditions. Under these conditions, rather than baseline features and classification methods, a neural network classifier using more advanced features can detect speech activity more successfully. Besides, it is seen that speaker diarization can be used in speech activity detection problem in a more advantageous way, and finally a two-stage speech activity detection system is designed. The system has a robust performance in real-life office recordings in which baseline speech activity detection systems perform poorly.