What It Does
AMPL detects cognitive pressure from voice alone — before any visible change in behavior. No wearables, no cameras, no self-report. Just the acoustic signal that the speaker’s motor system is already under strain.
The voice is a motor system. When someone is cognitively compromised — by stress, fatigue, overload, high stakes — their vocal production changes at the spectral level. These changes are invisible to listeners but measurable by instrument.
Why It Matters
Every high-stakes system assumes the human in the loop is thinking clearly. Pilots, surgeons, commanders, AI oversight operators — their cognitive state is the actual safety margin. Today, there is no instrument that monitors that margin in real time.
AMPL is that instrument. The sensor is already in everyone’s pocket.
What We Found
We initially predicted that pressure would freeze the vocal system (reduce spectral complexity). The data said the opposite: pressure causes activation — spectral complexity increases via sympathetic nervous system arousal. The hypothesis was falsified; the signal is real and stronger than expected.
Phase 0a: 847 speech segments from naturalistic conversation under cognitive pressure. Permutation test, FDR-corrected. p = 0.0001. Multi-method convergence across four independent detection approaches.
Phase 0b: Cross-dataset replication on an independent corpus. p = 0.012. Consistent effect direction — activation, not freezing.
Validation Roadmap
| Phase | Question | Status |
|---|---|---|
| P0a | Can acoustic features detect cognitive pressure at all? | Complete — p=0.0001 |
| P0b | Does the signal replicate across independent datasets? | Active — p=0.012 |
| P1 | Is the signature univariate or multivariate? | Upcoming |
| P2 | Does it replicate across speakers and languages? | Upcoming |
| P3 | Can it be induced and measured under controlled conditions? | Upcoming |
| P4 | Does it generalize to real-time detection? | Upcoming |
| P5 | Can external observers be reliably calibrated against the instrument? | Upcoming |