Window Size Versus Accuracy Experiments in Voice Activity Detectors

TL;DR

The paper investigates how analysis window size affects VAD accuracy across three detectors—RMS, WebRTC, and Silero—on a 1.1-hour diverse dataset, with additional evaluation of hysteresis post-processing. It uses metrics including ROC AUC, PR AUC, and MCC to capture performance under class imbalance, and it analyzes window sizes from 10 ms to 10 s. The main findings show that Silero significantly outperforms WebRTC and RMS, while larger averaged windows generally reduce accuracy; hysteresis provides a notable boost for WebRTC but limited benefit for Silero and RMS. The work offers practical guidance for configuring VADs in real-world pipelines and highlights how ground-truth labeling conventions influence observed window-size effects.

Abstract

Voice activity detection (VAD) plays a vital role in enabling applications such as speech recognition. We analyze the impact of window size on the accuracy of three VAD algorithms: Silero, WebRTC, and Root Mean Square (RMS) across a set of diverse real-world digital audio streams. We additionally explore the use of hysteresis on top of each VAD output. Our results offer practical references for optimizing VAD systems. Silero significantly outperforms WebRTC and RMS, and hysteresis provides a benefit for WebRTC.

Figures (15)

• Figure 1: ROC for 100ms windows.
• Figure 2: Precision-recall for 100ms windows.
• Figure 3: RMS ROC across window sizes.
• Figure 4: WebRTC ROC across window sizes.
• Figure 5: Silero ROC across window sizes.