ElevenLabs launches Eleven Realtime voice model, gradually rolling it out API
BIBIBI
AT A GLANCE
ElevenLabs has released the Eleven Realtime voice model, including Mini and Max versions, and is gradually making it available through API.
Article
ElevenLabs bundles voice Agents: supports 90 multiple languages and is compatible with the OpenAI interface
Dongcha Beating AI Flash News: ElevenLabs has released its real-time voice model, Eleven Realtime, including Mini and Max versions. Mini focuses on low-cost conversations, while Max targets customer service and complex tasks requiring continuous tool calls. Both support 90 multiple languages and 17000 voice options.
In Artificial Analysis's 1,000 voice reasoning questions, Max achieved an accuracy rate of 98.8%, and began speaking in an average of 0.57 seconds. Qwen Audio 3.0 Realtime Plus ranked slightly higher in accuracy at 99.2%, but took an average of 1.54 seconds to begin speaking.
Building a voice Agent typically requires independently connecting speech recognition, a large language model, and speech synthesis, as well as tuning pauses, turn-taking, and interruptions. Eleven Realtime optimizes these components together, eliminating the need to assemble them individually. Its underlying architecture still chains together multiple models; in Realtime mode, developers cannot select the individual models.
The new model has begun rolling out through ElevenAgents and Realtime API and is being gradually made available.API It is compatible with OpenAI's WebSocket protocol, allowing developers to reuse existing SDK; projects using WebRTC need to adjust their connection method first.
Original link https://m.theblockbeats.info/flash/373030
Key points
01
ElevenLabs has released the real-time voice model Eleven Realtime, including Mini and Max versions.
02
Mini focuses on low-cost conversations, while Max targets customer service and complex tasks requiring continuous tool calls.
03
Both versions support 90 multiple languages and 17000 voice options.
04
The article states that, among Artificial Analysis's 1,000 voice reasoning questions, Max had an accuracy rate of 98.8%, and began speaking in an average of 0.57 seconds.
05
The article states that Qwen Audio 3.0 Realtime Plus had an accuracy rate of 99.2%and took an average of 1.54 seconds to begin speaking.
06
Eleven Realtime jointly optimizes speech recognition, large language models, speech synthesis, and processes such as pauses, turn-taking, and interruptions.
07
Its underlying architecture still chains together multiple models; in Realtime mode, developers cannot select the individual models.
08
The new model has begun gradually rolling out through ElevenAgents and Realtime API.
09
API Compatible with OpenAI's WebSocket protocol, developers can reuse existing SDK; projects using WebRTC need to adjust their connection method.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is a product that bundles and optimizes the multiple components required to build a real-time voice assistant. Mini is oriented toward low-cost conversations, while Max is oriented toward customer service and complex tasks involving continuous tool calls. Developers can integrate through an interface compatible with OpenAI's WebSocket protocol, but cannot separately select the underlying models in Realtime mode.
Why it matters to readers
It may reduce the work developers must do to independently connect and tune components such as speech recognition, large language models, and speech synthesis, while also offering multiple languages and a wide selection of voices.
Beginners can understand it as a relatively complete real-time voice Agent solution; if they already use OpenAI's WebSocket protocol SDK, the article states that it can be reused, but WebRTC projects need to modify their connection method.
Risks and unknowns
The article does not include an official ElevenLabs announcement, so the specific regions of availability, pricing, and date of full availability cannot be confirmed.
The Artificial Analysis evaluation data does not include an original test link, so the test conditions and results cannot be independently verified based on the available evidence.
The specific scope and progress of the “gradual rollout” are not explained.
WebRTC The article does not provide specific technical details on how the project needs to adjust its connection method.
Related Developments
Loading event timeline…
Related concepts
voice Agent
This term is not in the glossary yet. Browse related concepts in the glossary.
