Towards a Digital Twin for Voice Pathology: Analyzing Microphone Effects on Acoustic Feature Extraction

Microphone variability is a potential source of bias in acoustic feature extraction, which is critical for applications such as automated voice pathology detection based on digital twins (DTs). In this study, we investigate the effect of different microphones on commonly used acoustic features, including fundamental frequency ( f 0 ) and its standard deviation, pitch difference, harmonics-to-noise ratio (HNR), jitter, shimmer, Shannon entropy, zero-crossing rate, spectral fatness, spectral roll-off, spectral contrast and signal skewness. We recorded ten subjects (five men and five women) sustaining the vowel /a:/ using 3 different microphones in controlled conditions. After removing silent parts, we segmented the signals into fixed-length windows for feature extraction. To assess the impact of microphone choice, we first evaluated the normality of each feature using the Shapiro–Wilk test. Given that many features did not follow a normal distribution, we employed the Kruskal–Wallis test to compare feature distributions across devices. If a significant effect was detected, pair-wise Wilcoxon tests with Bonferroni correction were conducted. Our results indicate that statistically significant differences were found across microphones for the extracted acoustic features. These findings suggest that microphone variability has a significant effect on feature extraction in controlled settings, pointing out possible difficulties in DTs construction for voice pathology detection.