ElevenLabs 发布 Eleven Realtime v1 Mini 与 v1 Max 语音智能体模型

Building a voice agent that sounds natural today requires a lot of decisions. You have to make choices on individual components like speech recognition, voice activity detection, turn-taking, the underlying LLM, and text to speech. Building this stack is a project in itself, and getting all the components to work well together takes more testing and tuning.

Eleven Realtime is a new class of Speech to Speech models that removes those hurdles. Each model is a complete voice agent, optimized for separate use cases. Both support more than 90 languages and can use any of the 17,000+ voices in the ElevenLabs library:

Today, we’re launching:

• Eleven Realtime v1 Mini for natural conversations at low cost. This is the best choice for AI assistants, teachers, and interactive characters, and will be sufficient for most use cases.
• Eleven Realtime v1 Max for complex tasks. This is ideal for customer support, lead qualification, and multi-step workflows, where the agent has to follow a process and reason through steps.

Why we built realtime models

You shouldn’t have to become a configuration expert to build a voice agent that sounds natural. Getting an agent to feel and sound right is about small judgment calls that require nuance such as how long to wait after a pause before answering, how to recover when the user interrupts the agent, and how voice, prompts, and a model all work together. In ElevenAgents, these are all settings you can adjust.

The same goes for developers already on a realtime API. Ours is compatible with OpenAI’s Realtime API, so trying Eleven Realtime simply means changing the endpoint and model name in the code you might already have.

Built on cascaded architecture

Eleven Realtime is built on a cascaded architecture of the components described above, tuned end to end. We call them Speech to Speech models because that is what you get with audio in, audio out, one model ID, and a system that can be measured on the same leaderboards as fused and duplex models.

Fused models initially set the bar for how natural a voice agent can feel, while a cascaded setup meant increased latency at every handoff and components never tuned against each other. Realtime removes that trade-off, and the tuning that used to be your project is already done.

Cascaded also has practical advantages in production, because every step produces text. You can see what the agent heard and what it decided to say, so you can tell a misheard word from a wrong decision. That same text is what lets ElevenAgents apply guardrails and PII redaction to a Realtime agent. 

The benchmarks below are the test of whether that trade-off is really gone. On conversational dynamics, both models score in the same band as fused models. On accuracy and speed together, no model on the Artificial Analysis leaderboard is both faster and more accurate than v1 Max.

How they perform

Artificial Analysis evaluated both models independently across three public benchmarks: Big Bench Audio for speech reasoning, Full-Duplex-Bench for conversational dynamics, and the Speech Agent Arena for human preference and task success.

Speech reasoning

Big Bench Audio asks 1,000 reasoning questions by voice and scores the spoken answer. Realtime v1 Max answered 98.8% questions correctly, and v1 Mini answered 75.9% correctly.

Speed

Big Bench Audio asks 1,000 reasoning questions by voice and scores the spoken answer. V1 Max starts speaking in 0.57 seconds on average, and v1 Mini in 0.62 seconds.

The Pareto frontier for accuracy and speed

Accuracy and speed are usually seen as a trade-off. The models that score highest on Big Bench Audio take the longest to answer, and the fastest ones give up a lot of accuracy. v1 Max excels in both of these metrics.

Conversational dynamics and human preferences

On Full-Duplex-Bench, which tests pauses, turn-taking, interruptions, and backchannels, v1 Mini scores 95.8% and v1 Max 94.1%. In the Speech Agent Arena, where people compare two models in blind live conversations, v1 Max scored 998 and v1 Mini 988. On tool-use tasks, v1 Max completed 87.1% and v1 Mini 60.3%, which is why we recommend Max for use cases with heavy tool calling.

Get started

In ElevenAgents (recommended)

Create an agent, or open an existing one, and select a Realtime model from the dropdown. From there you have everything ElevenAgents provides for running agents in production deployment. That includes phone, web, WhatsApp, and CCaaS channels, workflows, a knowledge base, guardrails, testing, and monitoring.

Check out the ElevenAgents docs for more details.

Through the Realtime API

If you already have an integration with OpenAI's Realtime API, the fastest way to try Eleven Realtime is to keep your code and change where it points. The Realtime API connects over WebSocket and follows the OpenAI Realtime specs, so OpenAI's SDK works as is.

The Realtime API connects over WebSocket only, so a WebRTC integration needs to move to WebSocket first. It also runs separately from ElevenAgents, so if you need channels, a knowledge base, or other platform features, build on ElevenAgents from the start.

The migration guide covers the steps and the differences from OpenAI's API.