Hume AI

EVI is a real-time voice model built to hold a conversation, not just trade turns. It measures the tune, rhythm, and timbre of your speech as you talk, then picks up on whether you sound calm, frustrated, or distracted and answers in kind. It knows when you've finished a thought and stays interruptible, so you can cut in and it picks the thread back up. EVI 4-mini speaks English, Japanese, Korean, Spanish, French, Portuguese, Italian, German, Russian, Hindi, and Arabic.

Octave is a speech-language model, which means it understands what the words mean before it speaks them. It decides when to whisper and when to shout, and it adapts pitch, tempo, and pronunciation to fit the line. You can build a voice from a description or clone your own, and Hume states that you keep ownership of the audio you generate. It targets narration for audiobooks, podcasts, and character work rather than robotic announcements.

The expression API turns voice and face into structured numbers. From speech it returns scores for hundreds of expressions and vocal qualities, each a probability between 0 and 1, measured natively in more than 50 languages. From video it scores facial expressions and visible actions for the faces it detects. Timestamped results link each score back to the exact moment in the recording, so you can find the part of a call that went sideways.

VoiceEQ is Hume's answer to the fact that voice systems pass their tests and still disappoint people. It takes your real conversation data, clusters it into use cases, and generates test suites with the scenarios and success criteria that matter to you. You then stress-test models and agents under realistic conditions, combine automated scores with human ratings, and turn the failures into a short list of things to fix.

Real World VoiceEQ is a public benchmark that scores voice models on what listeners actually experience. It covers more than 15 dimensions and 60 metrics, drawing on human judgments gathered during development, and separates speech generation, speech-to-speech, recognition, and understanding. For teams choosing a model, it's a way to compare systems on naturalness, timing, and context rather than on transcript accuracy alone.

For teams that need fuel for training, Hume offers annotated datasets covering conversational, multilingual, and vocal-burst data across more than 50 languages. The Prism pipeline cleans, segments, transcribes, and tags audio in your own cloud, filtering out poor-quality recordings and personal information before they reach your model. Both are aimed at research workflows where data quality decides the outcome.

The Human Feedback API runs structured evaluations with real, screened raters, returning ratings on naturalness, preference, and clarity within hours. Alongside it, Hume ships Python, TypeScript, and React SDKs plus playgrounds for EVI and Octave, so a developer can move from an idea to a working prototype without building the plumbing from scratch.