ElevenLabs Scribe v3: speech recognition that stops cutting callers off mid-sentence

What's going on

Voice AI company ElevenLabs, best known for synthetic voices, has launched two new speech-to-text models, which turn spoken language into text: Scribe v3 and Scribe v3 Live. Scribe v3 handles pre-recorded audio files, while Scribe v3 Live is built for real-time use with low latency, meaning as little delay as possible between speech and finished text. According to The Decoder, both are available now through the ElevenAPI programming interface and on the ElevenAgents agent platform. They're designed for messy situations: meetings where everyone talks over each other, or phone calls with background noise.

The models were unveiled around the ElevenLabs Summit in New York on November 11, 2026. According to the company's press release, it announced 13 new products and upgrades there in total. They include memory for voice agents, an automation layer, a meeting app and several partnerships.

What Scribe v3 is supposed to do better

ElevenLabs says Scribe v3 makes 27 percent fewer word errors than its predecessor, Scribe v2, across 102 languages. German is said to show especially large gains. Diarization, figuring out which sentence came from which person, has also improved, according to the company, particularly in larger group conversations. With Scribe v3 Live, speaker identification now works in real time for the first time.

Developers also get more control over the output:

• Custom vocabulary: You can give the models thousands of specialized terms up front, such as brand names or medications, so they don't mishear them.
• Plain-language formatting: You can describe in ordinary sentences how numbers or dates should appear in the transcript, without having to code rules.
• Multilingual mode: If someone switches languages mid-sentence, the model is supposed to keep up. Industry outlet CIOL cites India as an example, where customers often jump between English and regional languages during a call.
• Enterprise privacy: There are isolated servers in the EU and a zero-retention mode in which data is supposed to be deleted right after the request.

For healthcare, there's also Scribe v3 Medical, a variant trained on medical terminology. ElevenLabs CEO Mati Staniszewski says a similar model for financial services is in the works.

The problem of callers getting cut off

Scribe v3 Live aims to fix a classic annoyance with phone voice agents. Existing systems often use a fixed pause timer: once it's been quiet for a set time, the utterance counts as finished and the bot replies. If you pause to think or read out a customer number slowly, you get interrupted mid-sentence. Scribe v3 Live therefore doesn't just listen to the audio signal, it also analyzes what's being said. That's meant to let it tell whether someone is really done. A half-finished "My number is 4, 7, 1 …" then won't be treated as a complete answer.

This capability hasn't been independently verified yet. The measurements needed to do that don't exist, as the next section shows.

Independent tests: very accurate, but slow and pricey

According to The Decoder, benchmark platform Artificial Analysis has already tested Scribe v3. On raw transcription accuracy, the model is right at the top. It edges out models from Microsoft, Mistral, xAI, Google and OpenAI. Only a preview model from Alibaba does better.

The downside: on speed and price, ElevenLabs lands only mid-pack. The Microsoft and xAI models make marginally more errors but run several times faster and cost only about half as much. If you need to transcribe large volumes of audio quickly and cheaply, you're better off there.

The benchmark only measures word errors on recorded files, though. Speaker identification, rare languages and how Scribe v3 Live behaves in real time aren't included. CIOL also points out that the 27 percent figure comes from ElevenLabs' internal tests. How well the models perform day to day depends on audio quality, language combinations, specialized vocabulary and deployment conditions.

What else ElevenLabs announced

The new models are part of a bigger enterprise push. According to the press release, ElevenLabs released three new speech models in the past month: Scribe v3; the text-to-speech model Eleven v4, which is supposed to take a script's tone, pacing and context into account; and Eleven Realtime, a new class of speech-to-speech models for fast voice agents.

The summit brought more news:

• Interaction Engine (preview): The system decides which model handles which part of a conversation. According to Unite.ai, it also manages who speaks when and how interruptions are handled. Simple tasks like triaging requests go to a cheap model, while hard cases get a frontier model, one of the most capable and expensive language models. That's meant to save tokens, the units AI providers bill by. CIOL notes that ElevenLabs hasn't provided independent figures on the potential savings.
• Memory: Voice agents recognize logged-in returning customers and remember preferences and constraints. Per Unite.ai, someone who prefers Spanish and is only free after 5 p.m. won't have to say so again on the next call. The company's CRM stays the system of record for customer data.
• Routines: An automation layer for multi-step tasks before, during and after a call. Examples include pre-screening loan applications with a callback within minutes, appointment reminders for patients, or offering canceled slots to people on a waitlist.
• Architect: A built-in assistant that lets teams build and refine agents by voice or text. Every change is tested in a simulation first and only goes live after approval.
• ElevenAI: A work app that turns calls and in-person meetings into notes and to-dos, with built-in notification and consent features. Free early access has started. Dictation and meeting notes run on Scribe v3.
• Model customization: In the coming months, select enterprise customers will be able to work with the research team to customize speech models and selected open-source language models. They'll use fine-tuning, retraining on their own data, and reinforcement learning, learning from rewards for good results. Targets can include metrics like customer satisfaction (CSAT). According to CIOL, there's no pricing or general availability yet.

"We believe in a future where each of you has your own RL stack," Staniszewski said on stage.

Partners, Nvidia and Poland

ElevenLabs has launched a partner program with more than 550 partners in 44 countries, including Google Cloud, IBM, Deloitte, DXC and Genesys, according to Unite.ai. There's also an alliance with McKinsey and its AI arm, QuantumBlack. On a new marketplace, more than 20 partners offer ready-made industry agents for healthcare, finance, insurance, education and retail, among others. Buying through existing ElevenLabs contracts won't be possible until 2027.

ElevenLabs has also expanded its work with Nvidia. Nvidia's open model Nemotron 3.5 Super VL is now available in ElevenAgents. ElevenLabs also plans to add 15 megawatts of compute based on Nvidia's Vera Rubin systems, on top of its existing Blackwell Ultra systems. The two companies are also building a voice agent for Centralna e-Rejestracja, the central appointment system of Poland's e-health platform. It runs locally on government infrastructure and calls patients who only have a landline to remind them of appointments and help with rescheduling. ElevenLabs says the system handles 40 million appointments a year, 10 to 20 percent of which are missed. Nvidia Confidential Computing is meant to keep citizens' data and audio protected during processing. According to The Decoder, ElevenLabs is also working on local voice control for robots.

The bigger picture

On accuracy, Scribe v3 is demonstrably top tier. That matters where mistakes are costly: doctor's appointments, account numbers, contracts. But the independent test also shows you pay for it, in both money and speed. For bulk processing, Microsoft and xAI are the more pragmatic choice. The genuinely new part, Scribe v3 Live's context-based end-of-turn detection, is so far just a vendor claim. The improved speaker identification hasn't been independently tested either.

Strategically, ElevenLabs wants to be more than a voice provider. It sees itself as the "interaction layer" between people and AI, putting it up against call center software, meeting assistants and the big AI labs. Many of the announcements are previews or in early access, though. Memory and meeting recording also raise questions: what do the agents store, for how long, and who gets access? Whether the plan works out will only become clear once the products are widely deployed.