Objective: To assess the technical equivalence and scalability of remote app-based recordings against gold-standard laboratory conditions to identify reliable speech features suitable for longitudinal monitoring of speech impairment in Parkinson’s disease (PD).

Background: Acoustic speech markers can detect subtle PD-related changes, but their interpretability depends on recording quality and cross-setup stability. Remote smartphone sampling enables scalable monitoring but may introduce environmental noise and device distortion that bias acoustic features. Cross-setup reliability must be established before implementing digital speech markers in real-world settings.

Method: Eleven healthy controls (HC) and 24 people with PD were recorded with two sequential setups [table 1]: (i) ki:elements Mili app and (ii) soundproof booth with an external microphone. Tasks covered sustained vowels, oral diadochokinesis (DDK), reading, and free speech. Acoustic features were extracted using SIGMA, ki:elements’ proprietary speech processing pipeline. We first quantified recording quality/noise characteristics to contextualize downstream feature comparability. Cross-setup robustness of acoustic features was then evaluated using intraclass correlation coefficients (ICC).

Results: Most noise metrics were similar between setups [table 2], while spectral flatness was higher in app recordings, indicating more broadband noise. Fundamental frequency (F0) based measures showed consistently good-to-excellent agreement across all tasks (ICC ≈ .85–.96), supporting device-independent use [figure 1]. Perturbation, harmonic, cepstral, and spectral-energy features showed moderate, task-dependent agreement (ICC ≈ .50–.75), with highest stability in reading. This suggests connected speech may be preferable when deploying multi-feature panels remotely.

Conclusion: This study identified features with high cross-setup reliability to support selection of robust, meaningful speech biomarkers for remote monitoring and high-precision patient characterization. Smartphone app recordings can support scalable acoustic monitoring when analyses prioritize device-independent endpoints. Across speech domains, F0 measures were reliably comparable between app and booth recordings, whereas voice-quality and spectral/cepstral features were more sensitive to recording conditions and should be interpreted cautiously or calibrated when pooling modalities.

Table 1

Table 2

Figure 1

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/comparability-of-smartphone-app-and-booth-recordings-for-robust-acoustic-speech-monitoring-in-parkinsons-disease/