comparison
Speech models read the words. Expression APIs score a mood. Neither is a signal you can build on.
Frontier voice models are built to converse. Specialty expression APIs are built to label affect. Amplifier is built to measure, with a labeled clinical corpus underneath and an evidence tier on every number.
scroll the table sideways →
| Amplifier | Hume AI | audEERING | OpenAI gpt-live | Gemini Flash | Grok Voice | |
|---|---|---|---|---|---|---|
| What it returns | structured clinical signal readouts | emotional expression scores | affect & paralinguistic descriptors | speech + inferred tone | speech + inferred tone | speech + inferred tone |
| Clinically labeled training samples | 2,000,000 | not disclosed | not disclosed | not disclosed | not disclosed | not disclosed |
| Signs covered, with evidence tiers | 21, three tiers | 53 language + 48 voice/face expressions, no evidence tiers | 11 modules, no evidence tiers | none (steering signal, not scored) | none (steering signal, not scored) | none (steering signal, not scored) |
| Fine-tune on your own voice data | yes | no (dataset licensing only) | no | no | no | no |
| Dedicated hosted endpoint per client | yes | no (multi-tenant API) | no (SDK / on-device) | no | no | no |
| Per-signal explainability | features + confidence + tier | score only | score/descriptor only | none | none | none |
| Health condition categories | 9 | 0 | 0 (research use, not shipped) | 0 | 0 | 0 |
Vendor capabilities change; claims verified at time of publication.