HumeUnclaimed

Hume AI builds tools for measuring and improving voice AI. Its pages describe a model that answers questions about audio directly, without relying on transcripts or separate models for each task. Developers ask a question about a clip, such as the speaker's tone, whether a reply addresses the caller, or whether background music is present, and get back structured answers with confidence scores.

The company offers several products. The Audio Decision API returns these answers from one model call. Expression APIs measure vocal expression, timing and conversational dynamics. User Testing API recruits participants to talk with a voice agent and rate it on criteria such as naturalness, ease and trust. Human Feedback API gathers ratings from trained raters on how people experience a voice system.

The evaluation platform covers both voice models and voice agents. A team defines what quality means for a use case, such as tone, timing and turn-taking. It then runs automated stress-test scenarios across accents, noise and interruptions, adds ratings from trained human raters, and compares checkpoints to see gains, tradeoffs and regressions. The pages show an example where an agent's reply came after a long pause and did not acknowledge an urgent caller, and the team's fix was to acknowledge first.

Hume also publishes a benchmark called VoiceEQ, which assesses voice systems in real conditions and combines objective metrics with human insights. The page lists results for several voice models.

The company says it works with frontier AI labs and enterprises. Its solutions pages list healthcare, insurance, financial services, customer support, travel and hospitality, and telecom. The site also lists a speech data offering for preparing audio for training.

The pages do not state a price. Access appears to be through the API and by contacting the company, and the site links to API documentation.