EVI 2 debuts Voice Control to shape AI speech with sliders

If you have spent any time talking to modern AI voice assistants, you have probably noticed they all start to sound suspiciously alike. For product teams, building a voice that actually fits a brand has usually meant picking between generic presets or wading into the legal and technical swamp of voice cloning.

Emotional AI startup Hume AI wants to give developers a third option. The team just rolled out Voice Control, an experimental feature for its Empathic Voice Interface 2 that lets you sculpt unique AI voices using continuous sliders instead of text prompts or audio clones.

Dialing in tone like an audio engineer

Human voice perception is notoriously hard to describe in words. You can easily pick out a friend's voice in a loud room, but try explaining what that voice actually sounds like to someone else and words fail almost immediately.

To bypass that language bottleneck, Hume mapped out ten continuous perceptual dimensions that developers can tweak on top of base voices. The sliders cover masculine or feminine tilt, assertiveness, buoyancy, confidence, enthusiasm, nasality, relaxedness, smoothness, tepidity, and tightness.

Every slider moves between negative 100 and positive 100, sitting at zero by default. The technical challenge here was disentanglement. In real humans, vocal traits are deeply intertwined. Hume developed an unsupervised method designed to ensure that turning up nasality or enthusiasm doesn't accidentally warp unrelated vocal qualities.

Everything is available right now in a no-code playground. You pick a base voice, drag sliders, preview the result in real time, and deploy it straight into your setup. Hume did note that extreme parameter combos are still in beta and might occasionally take a hit on audio quality.

Why it matters

We think this is a remarkably smart way to tackle one of the biggest headaches in voice AI: brand identity without the baggage.

Voice cloning has become a massive liability for serious businesses, between copyright concerns, impersonation risks, and unpredictable training data. Hume's slider system completely sidesteps cloning by generating synthetic voices that belong to nobody in the real world, yet feel nuanced and distinctly tailored to a specific character.

For developers and creators, it is absolutely worth playing with right away. Text prompts have always been notoriously finicky for audio styling because a prompt that sounds warm in one turn might come off flat or sarcastic in the next. Numerical sliders bring engineering predictability back to voice design, making tones fully reproducible across sessions.

Zooming out, this shifts the battleground for real-time voice interfaces. Just having low latency and fluent speech is no longer enough to stand out. If Hume can smooth out the rough edges on extreme parameter combinations, competitors leaning solely on static voice libraries will find themselves looking outdated very quickly.