Spectral super-resolution for Parkinson's voice via representation-level methods under mixed-reality acquisition

Voice is a practical remote biomarker for Parkinson's disease (PD), but real-world capture often yields low-resolution time-frequency inputs that under-resolve diagnostically salient microstructure. In this controlled empirical comparison, we test whether spectrogram super-resolution (SR) at the feature level performed inside the model rather than via waveform resynthesis improves PD vs. healthy control (HC) discrimination under realistic constraints.

Speech was recorded with a Microsoft HoloLens 2 using a standardized mixed-reality (MR) protocol from 161 speakers (75 PD/86 HC) across five tasks: Task 1 image description, Task 2 question answering, Task 3 story repetition, Task 4 sustained vowels, and Task 5 word repetition (DDK). Raw audio was exported as 48 kHz, 16-bit PCM, downmixed to mono, amplitude-normalized, conservatively trimmed for leading/trailing silence, and resampled to 16 kHz. Log-mel spectrograms (80 bins) were fed to practical ImageNet-pretrained backbones (ConvNeXt-Tiny, ResNet-50, EfficientNetV2-S). We compared six super-resolution (SR) strategies: identity, nearest (deterministic), bilinear SR, kernel/CARAFE-like SR, LIIF-like SR, and a frozen universal feature SR module (AnyUp) that upsamples intermediate feature maps. Evaluation used 5-fold, speaker-disjoint cross-validation with AUROC (AUC) and accuracy (ACC).

AnyUp was the most consistent top performer, ranking first in 11/15 backbone-task cells. Gains were largest on tasks dominated by fine spectro-temporal cues: for ConvNeXt-Tiny, AnyUp vs. identity improved sustained vowels (Task 4) by ΔAUC 0.030/ΔACC 0.070 and DDK (Task 5) by ΔAUC 0.054/ΔACC 0.045. Macro-averaged over tasks, AnyUp outperformed identity by +0.025 AUC/+0.045 ACC (ConvNeXt-Tiny), +0.015/+0.027 (ResNet-50), and +0.052/+0.048 (EfficientNetV2-S). Representative best-in-class results include AUC/ACC of 0.899/0.886 (Task 4) and 0.927/0.897 (Task 5) for ConvNeXt-Tiny+AnyUp, and 0.940/0.903 (Task 5) for ResNet-50+AnyUp.

Densifying spectrogram representations with a frozen, universal feature super-resolution module yields consistent, compute-efficient improvements in PD voice classification under MR-standardized acquisition, with the largest benefits on sustained vowels and DDK. The contribution is empirical rather than architectural: we compare practical representation-level SR choices rather than introduce a new SR module. Feature-level super-resolution is therefore a pragmatic alternative or complement to bandwidth extension when waveform synthesis is unnecessary. At a macro level, the observed accuracy gains (e.g., +0.045 for ConvNeXt-Tiny) suggest that representation-level SR may be operationally useful in low-resolution clinical-audio settings.