ElevenLabs gives voice agents real emotion in 70+ languages

Getting an AI voice to sound natural is less about clarity and more about knowing when to laugh or whisper.

ElevenLabs just rolled out Eleven v3 Conversational, its most expressive real-time speech model yet, now generally available through ElevenAgents and ElevenAPI.

Directing AI with laughter and whispers

The standout feature here is audio tags. Developers can now fine-tune vocal delivery on the fly, instructing the model to whisper, laugh, or inject a dose of sarcasm right into the conversation.

The model supports over 70 languages, including German, Spanish, French, Portuguese, and Hindi, alongside a sprawling library of more than 11,000 voices.

Pricing kicks off at $0.05 per 1,000 characters, scaling down as usage grows.

Why it matters

Real-time voice agents usually suffer from one big flaw: they sound like they are reading bullet points. For anyone building interactive NPCs, language tutors, or customer support bots, adding realistic emotional cues makes conversations instantly more believable.

ElevenLabs is smart to focus on emotional range rather than just latency or generic fidelity. Giving developers direct prompt-level control over tone makes implementation simple.

That said, high-volume real-time audio still adds up fast. While five cents per thousand characters is accessible for testing, teams building heavy conversational workloads will want to keep a close eye on their API bills.