Why Voice Biomarkers?
Text alone misses critical information. Voice carries physiological signals that reveal what words don't say.
The Gap Between Words and Reality
In any voice conversation, there are two channels of information:
| Channel | What It Captures | Signal Reliability |
|---|---|---|
| Text | What someone says | Easily curated, minimized, or performative |
| Voice Biomarkers | How they say it | Physiological—reveals what words hide |
When these channels disagree, you learn something important.
Cross-Domain Examples
Contact Center: Frustrated but Polite
A customer says: "It's fine, I understand these things happen."
Text analysis: Positive sentiment, no escalation risk.
Voice biomarkers: Elevated stress (0.65), irritability (0.58), low happiness (0.12).
The customer is frustrated but being polite. Without voice biomarkers, this call gets marked as resolved. With them, you can proactively offer compensation or escalate to retention.
Education: Confident Words, Anxious Voice
A language student says: "Yeah, I think I understand. Let's try the next exercise."
Text analysis: Ready to proceed.
Voice biomarkers: Anxiety probability (0.62), fear (0.28), low confidence indicators.
The student is masking confusion to avoid embarrassment. A tutor acting on text alone moves forward; one with biomarker insight pauses to check understanding.
Mental Health: Minimization
A user says: "My mood has been so-so, ups and downs. But I tried to remain positive."
Text analysis: Mild concern at most.
Voice biomarkers: Depression probability (0.78), distress (0.81), elevated anhedonia.
The user is minimizing significant distress. This pattern is well-documented in clinical research: individuals in crisis frequently underreport symptoms.
Employee Wellness: Surface Acting
A tutor says: "No problem at all! Let's try it another way!" (bright, encouraging tone)
Text analysis: Positive, engaged.
Voice biomarkers: Burnout (0.62), stress (0.68), fatigue (0.60).
The employee is performing positivity while exhausted. This "surface acting" accelerates burnout. Early detection enables intervention before they quit.
Coaching: Hidden Resistance
A coaching client says: "That's a good idea. I'll definitely try that."
Text analysis: Agreement, commitment.
Voice biomarkers: Low engagement (0.25), elevated stress (0.45), flat affect.
The client is agreeing to end the conversation, not because they're convinced. A coach with this insight can probe deeper rather than assuming buy-in.
The Concordance Framework
Sentinel performs explicit concordance analysis between text and voice:
| Pattern | Text Signal | Voice Signal | Interpretation |
|---|---|---|---|
| Concordance | Positive | Calm/happy | Genuine positive state |
| Concordance | Negative | Distressed | Acknowledged difficulty |
| Minimization | Positive | Distressed | Masking or poor insight |
| Amplification | Negative | Calm | Venting without crisis |
This concordance signal is what enables Sentinel to reduce both false positives (flagging venting as crisis) and false negatives (missing masked distress).
Why This Matters for Voice AI
Voice AI systems are increasingly handling sensitive conversations: healthcare triage, customer support, coaching, education, mental health. These systems need to:
- Detect when users aren't saying what they mean — minimization, politeness, performance
- Avoid over-reacting to language patterns — "I'm dying of embarrassment" isn't a crisis
- Provide ground truth for ambiguous situations — when text alone is insufficient
Voice biomarkers provide the physiological layer that makes this possible.
Research Foundation
The biomarker models are clinical-grade, validated against gold-standard clinical assessments. See the Nature Portfolio publication for Apollo validation and the Interspeech 2025 paper for related research.
See Biomarkers for the full list of available biomarkers.