EventsThe 2nd International Conference on Bioengineering
Published
This submission belongs to the session S3. Biosignal Processing of the event The 2nd International Conference on Bioengineering
Published date
16 Nov, 2026
Citation
Emil Svoboda, Andrzej Marciniak, Alfredo Morales Pinzon, Tomáš Bořil, Noriaki Wada, Boian Alexandrov, Hiroto Hatabu, Charles R.G. Guttmann, Vladimir I. Valtchinov, Smartphone-versus-Headset Speech-Feature Pipelines: An Exploratory Concordance Study in Healthy Adults, in Proceedings of The 2nd International Conference on Bioengineering, Barcelona, 10 November–13 November 2026, MDPI: Basel, Switzerland
Share
Email
Facebook
Twitter
LinkedIn

Smartphone-versus-Headset Speech-Feature Pipelines: An Exploratory Concordance Study in Healthy Adults

Andrzej Marciniak 2
Tomáš Bořil 3
Noriaki Wada 4
Hiroto Hatabu 2
Charles R.G. Guttmann 2
1. Institute of Formal and Applied Linguistics, Faculty of Mathematics and Physics, Charles University, Prague, Czech Republic
2. Department of Radiology, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA
3. Institute of Phonetics, Faculty of Arts, Charles University, Prague, Czech Republic
4. Center for Evidence-Based Imaging, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA
5. Los Alamos National Lab, Los Alamos, NM, USA
6. Department of Biomedical Informatics, Harvard Medical School, Boston, MA, USA
Abstract

Introduction:

We aim to create and compare two digital-health pipelines for voice recording and speech analyses based on smartphones and professional equipment. We benchmark smartphone recording quality on speech features previously shown predictive of neurological impairment.

 

Methods:

A mobile app records users performing three speech tasks on the built-in microphone while simultaneously recording with professional equipment. Speech features are compared across each recording pair. Tasks include text reading, 20-second syllable repetition, and vowel sustainment. Bland-Altman visualization, Deming regression, correlation tests (Spearman, Intraclass) and Kolmogorov-Smirnov tests were utilized to assess concordance degree across devices.

Dot plots and standard deviation were used to assess longitudinal stability. Signal-to-Noise ratios (inter-speaker feature differences divided by within-speaker feature differences) were utilized to assess across-device reproducibility. Temporal stability was evaluated via relative size differences between speakers, measurement error, and test-retest reproducibility.

 

Results and Discussion:

20 recording sessions were collected from 3 speakers, with one speaker recording 6 sessions spanning 5 months to explore longitudinal stability. Twenty-three speech features served as a basis for assessing concordance between recording pairs. 19/23 (80%) features were found to be significantly correlated across recording methods; concordance is dependent on feature family. Furthermore, the smartphone’s audio enhancement seems to be affecting some features. While most features (14/23, 63%) seem to be longitudinally unstable, this finding is preliminary as it is based on a single speaker.

 

Conclusions:

A smartphone-based voice recording and speech analyses pipeline appears potentially feasible, however inter-device variability and possible audio enhancing algorithms deployed in smartphones need to be handled with care. The study is constrained by its very small sample size, therefore much-larger-N Case-Control studies are needed to validate and extend these preliminary findings to clinical settings and across specific neurological disease states.

Keywords
Digital health
Speech analysis
Smartphone recording
Neurological impairment
Voice biomarkers
Feature concordance
Longitudinal stability
Assessing the Potential of Deep Learning to Infer Sagittal Spinal Alignment from Coronal Radiographs