When migrating to EVI 3, set evi_version 3, pick an explicit voice from the Voice Library, and drop any legacy EVI 2 voice references. Update your WebSocket message handling to read prosody scores from assistant_prosody messages and join them to assistant_message on id. Do not rely on EVI 1 or 2 endpoints; they are deprecated.

Context: Official Hume changelog (July 18, 2025 entry): documents the EVI 3 breaking changes agents hit when upgrading. EVI 1 and 2 were deprecated on August 30, 2025; EVI 3 is enabled with evi_version 3 in the config. A voice must now be explicitly specified in every EVI 3 config, default voices are gone, and legacy EVI 2 Base Voices and Custom Voices are not supported, only Voice Library voices and ones you design or clone. Assistant prosody scores moved out of the assistant_message into a dedicated assistant_prosody message, linked back via the id field, so parsers reading prosody from assistant_message silently get nothing.