ElevenLabs Scribe v3 Live Teaches Voice Agents Not to Interrupt: 25% Fewer Errors

Imagine this: you call a clinic and dictate your insurance policy number. At the fourth digit you fall silent because you are checking your card. The bot decides you are done and cheerfully replies: "I didn't catch that, please repeat." Sound familiar? The problem isn't a dumb language model. Most voice agents rely on a rigid silence timer, and that is exactly what ElevenLabs is trying to replace.

What was released

The company introduced two speech recognition models. Scribe v3 is designed for finished audio files, while Scribe v3 Live works in real time with low latency. Both are available through the API and the ElevenAgents platform. They were built for "dirty" audio: in meetings people talk over each other, and on calls there is street noise.

According to ElevenLabs, Scribe v3 makes 25% fewer word errors than the previous version, across 102 languages. The company also notes notable progress in German. Diarization, i.e. determining who exactly is speaking, has become more reliable in large groups. In the Live version it works in real time for the first time.

How Live understands that you have finished

The old scheme is simple: silence longer than N milliseconds means the turn is over. Set N too low, and the bot interrupts people who are thinking or slowly dictating digits. Set it too high, and every answer arrives with an awkward delay.

Scribe v3 Live looks not only at the audio signal but also at the context of the conversation. If a person says "my number is four, eight, one..." and stops, the phrase is clearly unfinished, and the pause is not treated as the end of the turn. If they say "thanks, that's all," it is fine to respond. This example is ours: ElevenLabs does not disclose in detail how the mechanism works.

For a product manager this is the main part of the news. Interruptions are considered one of the most irritating problems of voice interfaces, and they determine whether a customer gets to a result or hangs up.

There are also nice touches for developers:

• you can load thousands of specialized terms into the model, such as brand or drug names, so it doesn't mangle them;
• number and date formats are set by rules written in plain language;
• the multilingual mode survives a language switch in the middle of a sentence;
• enterprise customers get isolated servers in the EU and a zero-retention mode, in which data is deleted immediately after the request.

What the independent benchmark showed

Artificial Analysis confirmed that Scribe v3 is among the most accurate models on the market. It outperformed solutions from Microsoft, Mistral, xAI, Google and OpenAI, with only an Alibaba preview model ahead. The price in the test was $3.33 per 1,000 minutes.

On speed and cost, ElevenLabs landed in the middle of the list. Models from Microsoft and xAI run several times faster, make only slightly more errors and cost roughly half as much. If you transcribe huge archives of recordings and care about price and speed, the choice is not obvious.

The main caveat is that the benchmark measures only word errors on recorded files. It did not test diarization, rare languages or the behavior of the Live version. So the key promise of the announcement, not interrupting the speaker, is so far confirmed only by the company itself.

The business around the models

Alongside the models, ElevenLabs is expanding its enterprise business. In partnership with Nvidia, it is embedding the open Nemotron 3.5 Super VL model into ElevenAgents, reserving additional compute capacity on Nvidia, and developing local voice control for robots. The company is also launching a phone agent for Poland's healthcare system. Separately, it announced an alliance with McKinsey, an expanded program for system integrators and its own marketplace, where Salesforce, IBM and Zendesk offer ready-made voice agents.

What to do with this

If you are building a voice agent for telephony, run Live on your own real recordings. Count not only the word error rate but also the frequency of interruptions, because that is exactly what public benchmarks do not measure. If your field is medicine, finance or any other with strict terminology, test the dictionary upload. If compliance with European requirements matters, look at the EU servers and zero-retention. And for mass transcription of archives, first compare price and speed: cheaper models may be the smarter choice at almost the same accuracy.