Hume AI

Hume AI is a company focused on the emotional and expressive side of voice. Its current platform is aimed at teams building voice AI, such as voice assistants, phone agents and conversational models, and it helps them measure how those systems perform in real conversations. Hume's position is that a task can technically succeed while the conversation still fails, because people judge voice AI on timing, tone, emotional expression, interruptions, accents and context as well as on words.

The evaluation platform

Hume's evaluation product, called VoiceEQ, is built around a loop of four stages: generate, stress-test, diagnose and validate. It analyses a team's conversation data to identify real use cases, then automatically builds evaluation suites with test cases, scenarios and success criteria. Systems are stress-tested under realistic conditions to find the conversations that break them. Failures are diagnosed to show why they happened, and fixes are validated in later runs, creating a continuous improvement cycle. The approach combines automated measurement with human judgement. Hume also publishes benchmarks showing how voice models perform on these real-world dimensions.

Expression APIs

Hume also offers Expression APIs that measure more than 600 dimensions of expression across over 50 languages. Developers can use them to analyse vocal and expressive signals in speech, for example to understand how a user sounds as well as what they say. Hume points to more than ten years of emotion-science research and a dataset of over 100 million speech samples behind its models.

Who it is for

The platform is built for frontier AI labs and enterprises shipping voice products. Hume says it is used by five of seven leading AI labs. Smaller developers may find the Expression APIs the easiest place to start, while the evaluation platform is aimed at organisations with production voice traffic and a dedicated team.

Things to consider

This is an enterprise and developer product, not a consumer app, and it is reached through the sales team or API documentation, not a self-serve editor. Measuring emotion and expression raises reasonable privacy and consent questions. If you process recordings of real users, check how data is stored and make sure your own users have agreed to that use. Inferred emotional signals are probabilistic, so they should guide improvements rather than drive high-stakes decisions about individuals.