Hume AI

Hume AI is unusual in the voice market because it sells the measurement problem first and the speech models second. The four-product structure is coherent: Data Solutions for collection, Kairos for simulation, Expression Measurement for real-time signal, Human Feedback for ground truth. For a team running model development, that's the shape of a real evaluation loop — generate scenarios, run them, score against humans, catch regressions, repeat. The claim that ratings come back "in hours, not days" is the core value proposition, because human eval is normally the bottleneck that forces teams onto automated judges of uncertain trustworthiness. Hume's SLM Judge leaderboard is a candid answer to that problem: instead of arguing that LLM judges are good enough, it publishes which judges track human ratings most closely, so you can pick one you can defend. On generation, Octave is positioned as a speech-language model rather than a TTS reader, and the roadmap has moved in the direction of operator control. February 2026 added supplemental LLM support (claude-opus-4-6, gpt-5.1, gpt-5.1-priority, gpt-5.2, gpt-5.2-priority) plus a ZERO prompt expansion mode for full system prompt control. April 2026 exposed per-config turn detection and interruption settings — end-of-turn silence between 500 and 3000ms, speech detection threshold, prefix padding, and minimum interruption duration. May 2026 added an experimental temperature parameter to TTS endpoints, higher for variation, lower for consistency. These are the knobs that separate a demo from a production voice agent, and they arrived in the last six months. Two honest constraints. First, pricing is usage-based underneath the seat price — every tier has an additional-characters rate and an additional EVI 3 rate, so a fixed monthly number is a starting point, not a ceiling. Second, Hume's own research position, argued in July 2026, is that emotional intelligence is a training-time property rather than a prompt-time one — which is a strong claim, but it also means you should not expect prompt engineering to close an expressiveness gap the model doesn't have. If raw fidelity is your only axis, compare ElevenLabs directly. If you need to know whether callers are expressing frustration and whether your model is getting better at handling it, Hume is built for exactly that question.