How Language-Independent Are Privacy-Preserving Depression Speech Features?

Motivation

Depression changes how people talk: longer pauses, slower speech, flatter pitch. Our system IHearYou (PerCom 2026) picks up these changes from ambient audio in a household without recognising any words. It only measures timing and pitch patterns, so it preserves privacy by design. On English recordings, it works. The open question is whether it still works in German.

Because IHearYou ignores words entirely, it also ignores most of what makes languages different. What remains (pause length, speech rate, pitch range) is driven by how the body produces sound, not by grammar or vocabulary, and depression slows and flattens speech regardless of language. So removing words for privacy might accidentally make the system portable across languages.

Research question

Do privacy-preserving speech features for depression detection transfer across languages better than conventional full-feature approaches?

What you will do

You will set up the IHearYou edge hardware in a controlled lab environment and collect German speech from healthy volunteers performing normal and simulated depressive speech tasks (HSG ethics approval required). From these recordings and from the English DAIC-WOZ corpus, you will extract three feature sets of increasing richness:

privacy-preserving prosodic features only (about 20 features),
full acoustic features including MFCCs (about 100 features),
deep speech embeddings via wav2vec 2.0.

You will then train depression classifiers on one language and test on the other, measuring the accuracy drop (transfer loss) for each feature set. The core test is whether the privacy-preserving set has the smallest transfer loss. Finally, you will rank individual features by how stable they are across languages, to identify which ones are portable and which break.

What you bring

Python programming (audio libraries, PyTorch or scikit-learn is a plus).
Interest in speech analysis, ubiquitous computing, or digital health.
Willingness to work with hardware (edge microphone devices); everything is provided.

Previous theses in this line

Audio-centered Approach for Building a Multimodal Predictive AI Agent to Detect Depressive Behaviors, Jonas Länzlinger (IHearYou).
Explainable Multimodal-based Depression Awareness at the Edge, Tibor Haller (CML).

References

[1] J. Länzlinger, K. Müller, B. Stiller, and B. Rodrigues. “IHearYou: Linking acoustic features to DSM-5 depressive behavior indicators.” arXiv:2511.14801, 2025.

[2] G. Kiss, M. G. Tulics, D. Sztahó, A. Esposito, and K. Vicsi. “Language independent detection possibilities of depression by speech.” In Recent Advances in Nonlinear Speech Processing, Smart Innovation, Systems and Technologies, vol. 48, pp. 103-114. Springer, 2016.

[3] F. Tao, D. Li, S. Tang, X. Ge, W. Ma, A. Esposito, and A. Vinciarelli. “Validating computational markers of depressive behavior: Cross-linguistic speech-based depression detection with neurophysiological validation.” arXiv:2604.01533, 2026.

[4] N. Cummins, J. Dineley, P. Conde, F. Matcham, S. Siddi, F. Lamers, et al. “Multilingual markers of depression in remotely collected speech samples: A preliminary analysis.” Journal of Affective Disorders, 341:128-136, 2023.

[5] N. Cummins et al. “A review of depression and suicide risk assessment using speech analysis.” Speech Communication, 71:10-49, 2015.

[6] J. Gratch et al. “The Distress Analysis Interview Corpus of human and computer interviews.” Proc. LREC, pp. 3123-3128, 2014.