Hume AI started from emotion science research and has grown into a platform for measuring and improving voice AI. The premise is that transcripts miss most of what matters in a spoken interaction. A task can succeed while the conversation fails, because humans judge voice systems on timing, tone, emotional expression, interruption, accent, context and adaptation. Hume measures all of it.
The centerpiece is an audio-native model that answers questions about audio directly, returning structured probabilistic judgments. Instead of orchestrating separate speech-to-text, speaker ID, language ID, emotion and AI-detection models and then aligning their outputs, you send the audio and ask questions: is this harassment, what is the speaker's tone, was the reply generated by AI, is there background music. This collapses a multi-model pipeline into a single call and keeps the answer in context.
Around that core sits an evaluation stack. VoiceEQ assesses voice systems in real-world conditions, merging objective metrics with human insight, and publishes comparisons across leading voice models. The User Testing API recruits participants and runs surveys on naturalness, ease and trust, while the Human Feedback API brings trained raters into the loop. Together these let teams develop a voice model, deploy an agent, and then verify with evidence what actually improved.