The Pulse
ElevenLabs Launches v4 Speech Models for Voice Agents
ElevenLabs launched v4 and v4 Turbo on September 28, adding expression controls and support for 90 languages. The Turbo model targets voice agents with lower latency, while v4 can clone a voice from 10 seconds of audio.
AI.info Team · Written with AI · fact-checked ·
ElevenLabs launched two speech-generation models on Monday, September 28: ElevenLabs v4 and v4 Turbo. The release pairs a model built for expressive audio with a lower-latency version aimed at voice agents, and raises the company’s stated language support from 70 to 90.
One model for performance, one for speed
The v4 generation uses a new architecture that gives users more control over expression and improves voice cloning, ElevenLabs says. V4 can create a voice clone from 10 seconds of audio. The model also better preserves voice identity across longer passages and uses surrounding text to shape how a line is delivered, the company says.
V4 expands on the inline tags ElevenLabs introduced with v3. Users can stack multiple tags to direct a sequence of vocal choices, giving the model instructions about how delivery should change from one line to the next. That approach puts more control inside the script rather than leaving expression to a single setting applied to an entire passage.
Turbo targets live agent conversations
The Turbo version is designed for applications where generated speech has to keep pace with a live exchange. It has lower latency, ElevenLabs says, and can begin producing audio as the language model behind an agent starts generating its response. The aim is to make spoken replies feel less like a separate step after text generation.
V4 can also handle confrontations, escalations and holds differently, the company says, allowing an agent’s delivery to reflect what is happening in a call. That feature is aimed at customer-service use, where a uniform tone can sound out of place when a conversation shifts from routine questions to a complaint or a request for help.
Language support reaches 90
The new models support 90 languages, up from 70 in the previous version. The largest quality gains, ElevenLabs reports, are in Japanese, Brazilian Portuguese, Mandarin and Cantonese. The wider language coverage extends the models’ intended uses beyond English-language voiceovers and into multilingual content and agent interactions.
Voice identity across longer passages and multi-step expression control also serve different needs: a narrator must stay consistent through a long script, while an agent needs to change tone in response to the conversation. The release combines those priorities in v4, then offers Turbo for situations where response speed matters more.
Voice models face a crowded field
ElevenLabs’ release comes as speech-model competition includes Cartesia, Deepgram, Fish Audio, Boson and WellSaid Labs, alongside voice products from Google and OpenAI. The company has also built a large enterprise calling business: TechCrunch reports that more than 55% of its business comes from large companies.
The split between v4 and Turbo defines the new release’s practical choice: expressive generation for produced audio, or a lower-latency model for agents. Turbo can start returning audio while the agent’s language model is still generating its answer, ElevenLabs says.