UK AI audio unicorn ElevenLabs launches Realtime models to simplify voice agent development
ElevenLabs released Mini and Max realtime voice models, unifying underlying components to lower development barriers, with third-party tests showing 98.8% reasoning accuracy for Max.
On October 20, UK-based AI audio company ElevenLabs launched the Eleven Realtime series, including v1 Mini for natural conversation and v1 Max for complex tasks. The core innovation lies in simplifying the development of Voice Agents: previously, developers had to separately select and tune three independent components—Automatic Speech Recognition (ASR), Large Language Models (LLM), and Text-to-Speech (TTS). The Realtime series encapsulates these underlying capabilities into a unified interface, allowing developers to choose between Mini or Max based on business scenarios and directly configure roles, knowledge bases, and tool-calling logic without manual component stitching.
According to independent tests by third-party AI evaluation firm Artificial Analysis, Eleven Realtime v1 Max achieved a 98.8% accuracy rate in the Big Bench Audio speech reasoning evaluation, with an average first-audio response time of 0.57 seconds. In this specific ranking, Max placed third, behind StepAudio 3 Realtime from StepFun (99.7%) and Qwen Audio 3.0 Realtime Plus from Alibaba (99.2%). Notably, as a reference metric, Qwen Audio 3.0 Realtime Plus has a first-audio response time of 1.54 seconds, highlighting ElevenLabs' competitiveness in low-latency interaction.
In the Speech Agent Arena evaluation involving real human conversations, Max demonstrated a tool-use task completion rate of 87.1%, significantly higher than Mini's 60.3%, indicating that Max is better suited for multi-step workflows requiring frequent external API calls. However, in the Full-Duplex-Bench evaluation, Mini scored slightly higher (95.8%) than Max (94.1%) in comprehensive metrics such as pause handling and interruption response, confirming the official positioning of Mini as a solution for 'low-cost natural conversation'.
Hands-on testing by Zhidongxi showed that using voice commands via the Architect assistant to modify Agent settings indeed reduced menu navigation, but Chinese speech output carried a noticeable foreign accent, with verbose responses and occasional lag. Technically, ElevenLabs continues to use a cascaded architecture, where independent ASR, LLM, and TTS modules work together. The company acknowledges that while this architecture facilitates individual upgrades and error troubleshooting, the conversion from speech to text loses some tonal and emotional information, which requires additional modeling to compensate. Currently, both models are accessible via the ElevenAgents platform or APIs compatible with the OpenAI Realtime specification, priced by connection duration and token usage.