ElevenLabs v4 Pushes Voice Agents Toward Real Conversation

ElevenLabs is raising the ceiling for voice agents. The company launched ElevenLabs v4 and v4 Turbo on September 28, adding richer expression controls, lower latency, and support for more than 90 languages. The release targets a problem that has become increasingly visible in production: a voice agent can be technically correct and still feel painfully slow, flat, or unable to handle a difficult conversation.

What changed

According to TechCrunch, the new models can begin generating audio while the underlying language model is still producing its answer. That streaming path should reduce dead air in calls and make interruptions feel more natural. ElevenLabs also says v4 can clone a voice from 10 seconds of audio, preserve identity over longer passages, and follow stacked inline expression controls. The company has expanded language support from 70 to more than 90, with notable quality gains in Japanese, Brazilian Portuguese, Mandarin, and Cantonese.

Why it matters for builders

The practical shift is not just better narration. Voice agents increasingly need to manage holds, escalations, objections, and handoffs without sounding like they have fallen out of the workflow. Lower latency gives orchestration layers more room to stream partial responses, while richer prosody can signal whether an agent is clarifying, apologizing, or escalating. For teams connecting telephony, CRM, and automation tools, that means voice quality becomes part of system design rather than a final presentation layer.

ElevenLabs says more than 55% of its business now comes from large companies, while competition is intensifying from Cartesia, Deepgram, Google, OpenAI, and other speech startups. The next differentiator will be operational: consistent identity, safe cloning, reliable turn-taking, and measurable outcomes across real calls.