Hume AI plans to release its new EVI 3 speech-to-speech model on Thursday, promising what early users describe as some of the most lifelike voice cloning yet available. The system can generate a convincing replica of a speaker’s voice from as little as a few seconds of recorded audio, maintaining the original rhythm, tone and expressive pauses. Testers who gained early access said the model reproduced their own voices with nuanced inflection and emotional range, and could even mimic public figures such as the late astronomer Carl Sagan. Several users reported successful clones after providing samples of 15–40 seconds, far less than the minutes of training data typically required by earlier tools. EVI 3 also supports connections to external large-language models, including offerings from Groq and Anthropic, allowing developers to pair realistic speech output with sophisticated text generation. Hume AI, which has focused on bringing emotional intelligence to conversational systems, has not yet disclosed pricing or commercial terms for the service.