ElevenLabs targets AI assistants and customer support with two new voice models

Voice artificial intelligence (AI) company ElevenLabs has launched two speech-to-speech AI models that combine the systems to run a voice agent, targeting developers building conversational assistants and businesses automating customer support.

The company said its models – Eleven Realtime v1 Mini and v1 Max – support more than 90 languages and can use over 17,000 voices from their library.

ElevenLabs positions Mini as a lower-cost option for assistants, teaching and interactive characters, while Max targets AI agents that must follow certain steps to complete a task.


Speech-to-speech models accept spoken input and generate spoken replies, bringing the process of understanding a caller, producing an answer, and managing the exchange into one system.

Building voice agents typically requires developers to connect separate systems for recognising speech, detecting when someone is talking, deciding when to respond, generating an answer and converting that into speech, the company said.

“Building this stack is a project in itself, and getting all the components to work well together takes more testing and tuning,” ElevenLabs said. “You shouldn’t have to become a configuration expert to build a voice agent that sounds natural.”

ElevenLabs said Artificial Analysis, an independent evaluation platform that tracks and compares artificial intelligence models, evaluated both models across three public benchmarks covering spoken reasoning, conversation handling, human preferences and task completion.

On Big Bench Audio, which asks 1,000 reasoning questions by voice and scores spoken answers, Max answered 98.8% correctly, compared with Mini’s 75.9%, according to results the company published.

Max also began speaking sooner in that test, averaging 0.57 seconds compared with Mini’s 0.62 seconds, giving it a higher reasoning score.

On tasks that involved using software tools, Max completed 87.1% against Mini’s 60.3%, the company said, citing the difference as the reason it recommends Max for workflows that rely heavily on such tools. These include customer support and assessing prospective customers, where an agent needs to follow a process and reason through individual steps, according to the company’s description of the model.

Mini scored higher on Full-Duplex-Bench, which tests pauses, interruptions, turn-taking and brief acknowledgements during conversations, recording 95.8% against Max’s 94.1%.

In the Speech Agent Arena, where people compare two models through live conversations without knowing their identities, Max scored 998 and Mini 988, respectively, the company said.

ElevenLabs said its developer interface is compatible with OpenAI’s Realtime API, allowing developers to try the models by changing the connection address and model name in existing code.

However, the company said the interface currently supports only WebSocket connections, so developers using WebRTC, another method of connecting applications for live communication, must change their connection method first.