ElevenLabs 发布 Eleven Realtime 实时语音模型，语音推理准确率达 98.8%
ElevenLabs Launches Real-Time Voice AI Models With Up to 98.8% Speech Reasoning Accuracy
ElevenLabs 推出 Eleven Realtime 语音到语音模型系列，包含面向 AI 助手、教师等场景的 v1 Mini 和面向客服、线索筛选及多步骤流程的 v1 Max 两个版本。两款模型支持 90 多种语言和超过 17000 种音色，将语音识别、轮次切换、语音活动检测与语音生成整合为一个系统。在 Artificial Analysis 的评测中，Max 在 Big Bench Audio 上正确回答 98.8% 的问题，Mini 为 75.9%；Max 首音频平均耗时 0.57 秒，Mini 为 0.62 秒；工具调用任务上 Max 完成 87.1%，Mini 为 60.3%。Eleven Realtime 遵循 OpenAI Realtime 规范，开发者只需修改 API 端点和模型名即可迁移，但使用 WebRTC 的集成需改用 WebSockets。
推荐理由
AI 评分 69
原文给出两款语音模型的基准差异与 OpenAI Realtime API 兼容路径，便于开发者权衡成本与能力。
主题
正文
当前显示原文，中文全文译文尚未就绪。
ElevenLabs has launched Eleven Realtime, a family of speech-to-speech models for building voice agents, with two variants for different workloads, as the company looks to simplify the development of conversational AI systems.
Eleven Realtime v1 Mini targets applications such as AI assistants, teachers, and interactive characters, while Eleven Realtime v1 Max targets customer support, lead qualification, and workflows involving multiple steps.
The two models give developers a choice between cost and capability requirements. Mini is positioned for applications where conversational interaction is the primary requirement, while Max targets tasks that demand stronger reasoning and tool use.
Both models support more than 90 languages and can use any of the more than 17,000 voices in the ElevenLabs library. The models combine speech recognition, turn-taking, voice activity detection, and speech generation into a single system.
“You shouldn’t have to become a configuration expert to build a voice agent that sounds natural,” the company said in its announcement.
ElevenLabs said the models address challenges such as determining when to respond after a pause, handling interruptions and coordinating voice output with the underlying model.
In independent evaluations by Artificial Analysis across three public benchmarks, Eleven Realtime v1 Max outperformed Mini on speech reasoning and tool-use tasks.
On Big Bench Audio, which evaluates reasoning through spoken questions and answers, Max answered 98.8% of questions correctly, compared with 75.9% for Mini.
Max also recorded an average time to first audio of 0.57 seconds, while Mini took 0.62 seconds, suggesting higher accuracy.
However, Mini scored higher on Full-Duplex-Bench, which evaluates conversational dynamics, including pauses, interruptions, turn-taking and backchannels. It scored 95.8%, compared with 94.1% for Max.
In the Speech Agent Arena, which compares models through blind live conversations based on human preferences and task success, Max scored 998, while Mini scored 988.
The gap was wider on tool-use tasks. Max completed 87.1% of these tasks, compared with 60.3% for Mini. ElevenLabs recommends Max for applications that rely heavily on tool calling and multi-step workflows.
Compatible with OpenAI’s Realtime API
ElevenLabs is also targeting developers already building voice agents using OpenAI's Realtime API. Eleven Realtime follows OpenAI’s Realtime specifications, allowing developers to switch to ElevenLabs by changing the API endpoint and model name in their existing integrations.
The API connects through WebSockets and supports OpenAI’s SDK. However, developers using WebRTC integrations will need to move to WebSockets to use the service.
Developers can also access the models through ElevenAgents, the company’s agent platform, which provides deployment across phone, web, WhatsApp, and contact centre channels. The platform includes workflows, knowledge bases, guardrails, testing, and monitoring tools.