Restoring Over-Attenuated Sustained Vowel Phonations in Smartphone Calls of COPD Patients Using U-Net and Speech Foundation Models.

Voice data collected through smartphone-based telemedicine may serve as digital biomarkers, but sustained vowel phonation recorded during smartphone calls is often excessively attenuated, limiting its clinical utility. We aimed to characterize this codec-related over-attenuation and develop a restoration framework for attenuated sustained vowel phonation. We retrospectively analyzed 1553 face-to-face sustained vowel phonation samples from 288 outpatients with and without chronic obstructive pulmonary disease and generated paired training data by synthetically degrading clean recordings with time-varying attenuation patterns resembling real calls. We evaluated a 1D U-Net, U-Net with a Whisper encoder, and U-Net with a Wav2Vec 2.0 encoder using a composite loss combining mean squared error, multi-resolution short-time Fourier transform loss, and scale-invariant signal-to-distortion ratio (SI-SDR) loss. General-purpose speech restoration models performed poorly, whereas all proposed models achieved positive SI-SDR values. For full waveform restoration, the U-Net with Whisper achieved the highest mean SI-SDR (13.677 dB), significantly outperforming the U-Net alone (13.190 dB; P = 0.001). In attenuated regions, both foundation model-based architectures significantly improved masked SI-SDR versus the U-Net (both P = 0.001). Qualitative evaluation on real-call samples showed recovery of previously suppressed segments. These findings suggest that speech foundation model-augmented U-Nets can effectively restore codec-attenuated sustained vowel phonation and support more reliable remote voice biomarker collection.