Abstract
BACKGROUND Vocal fatigue is a common functional voice complaint among occupational voice users and is typically assessed using self-report and clinical evaluation in controlled settings. These approaches do not capture dynamic fatigue-related voice changes that occur during everyday voice use. Advances in mobile health (mHealth) and machine learning provide opportunities for smartphone-based voice monitoring; however, environmental noise remains a major challenge for real-world deployment. OBJECTIVE This study aimed to evaluate the technical feasibility and noise robustness of a smartphone-compatible system for detecting vocal fatigue–related voice changes under controlled but ecologically relevant acoustic conditions. METHODS A proof-of-concept system integrating voice isolation, acoustic feature extraction, and machine learning–based classification was developed for smartphone deployment. Voice recordings were obtained from a single speaker under fatigued and nonfatigued conditions and augmented with 12 real-world background noise recordings from 10 everyday environments at signal-to-noise ratios (SNRs) ranging from −10 dB to 30 dB. System performance was evaluated at the subsystem and end-to-end levels, including noise attenuation, feature extraction accuracy relative to reference software, classification accuracy across SNR conditions, and computational feasibility on mobile-compatible hardware. RESULTS Within this single-speaker proof-of-concept dataset, the complete system achieved an overall vocal fatigue classification accuracy of 85.81%, with balanced classification performance metrics, including precision (0.856), recall/sensitivity (0.861), and specificity (0.855). Classification accuracy remained above 80% at SNRs of −5 dB and higher, decreasing to 78.67% at −10 dB SNR. The voice isolation module achieved an average noise attenuation of 56.62 dB across SNR conditions, with consistent suppression within the human vocal range. Acoustic feature extraction demonstrated high agreement with reference software, achieving average accuracy above 88% at −10 dB SNR and above 90% at higher SNRs. In a smartphone-server configuration, in which audio recorded on the smartphone was processed on a cloud server, end-to-end processing took 8.38 seconds per sample (excluding data transmission), indicating feasibility for short-latency periodic mobile monitoring. CONCLUSIONS This proof-of-concept study demonstrates the technical feasibility of smartphone-based detection of speaker-specific fatigue-related voice changes under a range of noisy conditions. While performance declined under extreme noise, results support the potential of mHealth voice monitoring systems for tracking vocal fatigue outside the clinic. Further validation across multiple speakers and with clinically anchored outcome measures is required prior to clinical application. CLINICALTRIAL N/a

This publication has 22 references indexed in Scilit: