Hume AI unveils EVI 3 for custom voice personalities

Most voice assistants still talk to you like an over-caffeinated customer service rep who cannot read the room. Hume AI wants to kill that robotic flatline for good.

The research lab just unveiled EVI 3, its third-generation empathic voice model. Rather than chaining together separate transcription, reasoning, and text-to-speech tools, EVI 3 is a unified speech-to-speech foundation model where a single system hears, thinks, and speaks natively in audio.

Prompting an AI to sound out-of-breath

Instead of restricting users to a handful of canned personas, EVI 3 lets you prompt any voice and personality into existence on the fly. It can tap into more than 100,000 custom voices already built on Hume's platform, adjusting styles implicitly based on context or explicitly by command.

In testing, Hume pushed the model through 30 distinct vocal tones and attitudes, ranging from whispering and breathless sprinting to pedantic know-it-alls and full-blown pirate impressions. Crucially, it works both ways: in blind tests evaluating how well the AI picked up on tone across nine distinct human emotions, users consistently rated EVI 3 higher than OpenAI's GPT-4o for both detection accuracy and conversational empathy.

On the technical side, Hume tackled latency by building a streaming architecture trained with reinforcement learning. While hardware-level latency can dip below 300 milliseconds, real-world tests from Hume's New York team logged an average delay of 1.2 seconds, well ahead of GPT-4o Realtime API's 2.6-second average and neck-and-neck with rival voice engines. It also operates on a fast-and-slow thinking pattern, allowing the voice to stream immediately while querying reasoning models and live web search in the background.

EVI 3 is available right now via a live web demo and Hume's iOS app, with API access rolling out to developers over the coming weeks.

Why it matters

We think EVI 3 shows that voice AI is finally graduating from utility to genuine social presence.

For builders creating virtual companions, interactive NPC gaming characters, or high-touch coaching tools, this is an immediate green light to experiment. Lowering the bar from expensive voice cloning pipelines down to a simple text prompt completely rewrites the economics of building unique digital characters. However, if your use case is dry, mission-critical workflow automation, all that performative warmth might just get in the way.

For the broader industry, Hume is mounting a serious threat to generic frontier labs. OpenAI and Google have treated expressive speech as a downstream showcase for their massive multimodal LLMs. Hume is proving that treating prosody, rhythm, and emotional nuance as first-class citizens creates a noticeably more engaging product.

The obvious tension lies in trust. As voice models get exceptionally good at acting distressed, joyful, or conspiratorial, the psychological surface area for manipulation expands drastically. An AI that can cry convincingly on demand is an incredible technical achievement, but it will inevitably test our collective ability to detect deception.