Automatic Silence Detection Employing Artificial Intelligence for Clinical Context Analyses

Automated speech and pause/silence detection is a crucial task in clinical and pathological environments, supporting diagnostic processes and providing essential information for treatment planning. This study evaluates three methods for automatic silence detection in clinical speech analysis: (1) a traditional energy-based method using zero-crossing detection, (2) a pretrained neural network model for voice activity detection (Silero-VAD), and (3) NVIDIA's speaker diarization and transcription tool. All methods demonstrated effective pause/silence detection with comparable error rates, though Silero-VAD exhibited superior precision and performance. Key metrics included a Dice coefficient of 0.917, an onset error of 500 ms, and an endpoint error of 370 ms, highlighting the importance of audio preprocessing.