The Startup Using Voice to Detect Depression, Fatigue and Diabetes

Preview clips featuring key insights about voice biomarkers, 15-second assessments, driver fatigue detection, and diabetes detection through voice.

Hermes welcomes Emilia Molimpakis from Thymia. Emilia shares the deeply personal story that inspired Thymia - her best friend's depression and suicide attempt, which led her to quit her postdoc and start a company using voice as a biomarker for health.

Emilia explains the three layers of speech analysis: acoustics (sound wave properties), content (words and meaning), and timing (pauses, breathing). How these connect to mental health and the entire nervous system, including depression, anxiety, respiratory health, cardiovascular health, and metabolic health.

The challenges of building a health biomarker dataset - why scraping the internet doesn't work. The importance of human-curated data, longitudinal collection, and multimodal data (video, behavioral, EEG, blood biomarkers). How gamification helped engage people with depression.

Voice's ubiquity and ease of access make it ideal for health monitoring. The challenge of isolating health signals from demographic signals (age, accent, ethnicity). How Thymia differentiates condition-specific signatures from personal characteristics.

Thymia's multilingual capabilities across Greek, English, Spanish, Italian, Brazilian Portuguese, Indonesian, and Japanese. The breakthrough of building individual longitudinal models where each person becomes their own baseline over time.

The evolution from 10-minute assessments to just 15 seconds with real-time processing (0.2 seconds). Hermes shares his experience trying the demo at a train station and being amazed by the accuracy.

Thymia's applications in healthcare (NHS, pharma companies) and safety-critical environments. The expansion from mental health to detecting fatigue and stress in automotive, aerospace, construction, mining, and defense industries.

How Thymia is creating a health and safety infrastructure layer for voice-first communication. Enabling safe voice agent deployment at scale by detecting distress, stress, depression, panic, and pain in real-time voice interactions.

How Thymia gained trust through rigorous science - 290+ peer-reviewed publications from the team, certifications, and medical device regulation. The trust domino effect from healthcare to safety-critical industries.

The surprising discovery that diabetes can be detected through voice. How chronic hyperglycemia affects vocal cords, lung capacity, and nerve coordination, creating detectable signatures. The rigorous validation process to ensure it's actually diabetes, not correlated factors.

Thymia's validation process: hypothesis, data gathering, multimodal controls, synthetic data generation, model building, productionization, and peer-reviewed publication. Papers currently under review at Nature and other major journals.

The huge difference between academic research and real-world deployment. Unpredictable human behavior (like users wearing towels over their heads), environmental variability, privacy concerns, and the need to account for background noise, multiple voices, and different devices.

Novel effects discovered in voice: how people compensate for loud environments, the need for different models in cars vs. other environments, and handling multiple voices with proper consent and voice locking.

How assumptions changed: from needing multimodal data in deployment to voice-only, from 10-minute samples to 15 seconds, from minutes of processing to 0.2 seconds. This speed enabled completely new applications like real-time health and safety infrastructure.

2025 as the year of voice. Upcoming milestones: deployment at scale with telecoms and LLM providers, regulatory approval as Class II medical device, diabetes clinical trials comparing against blood biomarkers, respiratory and cardiovascular health models, and edge deployment.

The biggest challenge: regulators initially forcing companies to regulate specific model versions, which stifles AI iteration. How Thymia helped change this to regulate inputs/outputs instead, allowing model updates while maintaining regulatory compliance. Speaking at EU Parliament to advocate for change.

Common investor misunderstandings: underestimating the difficulty of building health biomarker models, thinking big companies can do it overnight, and asking 'what models do you use?' when Thymia built everything from scratch. The value of human-curated datasets vs. scraping the internet.

The overemphasis on automating existing processes vs. deeper AI understanding. Example: automating broken mental health questionnaires doesn't solve problems - it creates bottlenecks. Thymia's approach: using biomarkers for intelligent triage and differentiation.

The importance of clinician-in-the-loop. Thymia can detect symptoms like appetite problems but can't determine root causes - doctors need to ask the right questions. AI augments human capabilities but doesn't replace human empathy and judgment.

If not working with voice, Emilia would explore consciousness itself - understanding what human consciousness is and whether it can be replicated in AI. Closing remarks and thanks.