Research on Voice Activity Detection Methods Based on Deep Learning

Voice Activity Detection (VAD), as a crucial component of the speech processing, distinguishes between speech and non-speech segments within a voice. By accurately identifying moments of speech, it enhances the efficiency and performance of speech processing, reducing the wastage of resources on non-speech segments. This paper introduces a deep learning-based end-to-end trained VAD model that ingests Log-Mel features and combines Convolutional Neural Networks (CNN) with Bidirectional Long Short-Term Memory networks (BiLSTM), incorporating an attention mechanism to refine the selection and extraction of speech features. We compared three baseline models proposed on the AVA-Speech dataset and validated the enhancement in model performance due to the chosen sequence data processing network and the integration of the attention module through ablation studies. Results on the AVA-Speech dataset demonstrate that our method achieves an ACC of 90% and an AUC of 0.9439, outperforming other models and effectively fulfilling the target task.