ElevenLabs CEO Mati Staniszewski predicts voice will replace text as the primary interface for AI, driven by low-latency and emotive models.
Why voice is moving ahead of text
ElevenLabs CEO Mati Staniszewski argues that voice will become the primary way people interact with AI, not a secondary feature bolted onto chat. The case rests on two product shifts already underway: models that respond fast enough for conversation, and models that carry tone, emphasis, and emotion instead of flat, synthetic speech. When both are good enough, talking to software stops feeling like a demo and starts feeling like a normal way to get work done.
Text remains excellent for precision, search, and audit trails. Voice wins when the task is continuous, hands-busy, or social: drafting while walking, coaching a teammate through a process, reviewing a document out loud, or navigating a product without hunting through menus. The interface question is not “voice or text forever,” but which channel carries the main loop while the other supports capture, editing, and reference.
What low latency actually buys you
Conversation collapses if the system pauses too long after you stop speaking. Low-latency voice models shrink that gap so turn-taking feels human: you interrupt, correct yourself, and keep momentum without restarting the whole request. That changes how people prompt. Instead of packing every constraint into one careful paragraph, they iterate out loud—narrow the scope, change the audience, ask for a shorter version—because the cost of another turn is low.
For builders, latency is a product budget, not just a model score. End-to-end delay includes speech recognition, reasoning, and speech synthesis, plus network and UI overhead. If any stage is slow, the conversation breaks even when the underlying model is strong. Practical systems stream partial results, allow barge-in, and fall back to text when silence or noise would otherwise stall the user.
Emotive models and when they matter
Emotive models make delivery part of the answer: calm for support, crisp for instructions, warmer for coaching. That is not decoration. People infer trust, urgency, and clarity from tone as much as from wording. A correct answer delivered flatly can still feel wrong; a slightly imperfect answer delivered clearly often works better in real use.
Use emotion with restraint. Customer-facing agents, training, accessibility tools, and consumer assistants benefit most. Internal tools that produce logs, code, or compliance records still need neutral, reviewable text as the system of record. Design for both: speak the result, then leave a transcript the user can edit, share, or store.
- Prefer voice for multi-step tasks where hands or eyes are occupied.
- Keep text for confirmation, citations, and anything that must be exact.
- Stream audio early; do not wait for a full monologue before the first word.
- Expose a transcript so users can correct names, numbers, and constraints.
How to design products if voice is the default
If Staniszewski’s prediction holds, product teams should stop treating voice as a microphone button on a chat box. Start from spoken goals: “book this,” “explain that,” “change the draft.” Structure flows around short turns, explicit confirmation for irreversible actions, and easy recovery when the model mishears a name or number. Multimodal still matters—screen, keyboard, and voice together—but the primary path should be speakable without a style guide.
Measure success the way conversation works: time to first useful utterance, recovery rate after errors, and whether users finish the task without switching back to typing. Voice becomes the next interface for AI only when those metrics beat text for the jobs users actually care about, not when demos sound impressive in isolation.