Hume AI EVI 3

Hume AI EVI 3 is the third generation of the Empathic Voice Interface, released as a developer API in July 2025. It is a native speech-to-speech foundation model that detects user emotional states from prosody and vocal patterns in real time, then adjusts its own tone, pace, and affect to respond empathically, without requiring any fine-tuning or external emotion classifiers. EVI 3 supports 30 distinct emotions and speaking styles triggered by natural language instructions or system prompts. Voice cloning is built in, requiring under 30 seconds of reference audio to capture timbre, accent, rhythm, and personality. The model integrates with third-party LLMs including Claude 4, GPT, and Gemini through a BYO-LLM configuration, and developers can access over 100,000 community-designed voices through the Hume voice library. EVI 3 is purpose-built for companion applications, mental health tools, wellness coaching, and consumer voice experiences where emotional resonance defines the product. Its expression measurement model outperformed GPT-4o on 8 of 9 tested emotional dimensions in independent evaluations, making it the strongest choice for voice products where tone and affect are core product features.