ElevenLabs launches Scribe v3 and Scribe v3 Live transcription models

By Elena Vasquez

ElevenLabs says Scribe v3 cuts word errors by about 25 percent across 102 languages, but Artificial Analysis found the model slower and pricier than key rivals.

A caller pauses to think, or slowly reads out a string of numbers. According to ElevenLabs, a rigid voice-agent timer can mistake that silence for the end of a turn, then cut the person off mid-sentence. The company designed Scribe v3 Live to address that specific failure.

The wider release covers two distinct jobs. Scribe v3 handles pre-recorded audio, while Scribe v3 Live targets low-latency, real-time applications. Both models are available now through the ElevenAPI developer interface and the ElevenAgents platform, according to ElevenLabs.

What changed in ElevenLabs' speech stack?

The models target audio where people talk over each other in meetings or background noise fills a phone call, according to ElevenLabs. The release covers both recorded conversations and live calls. It gives no model sizes, latency figures, or price points for either model.

Across 102 languages, ElevenLabs reports about 25 percent fewer word errors than the predecessor. German accuracy improves significantly, according to the company, though the release gives no separate German score. Artificial Analysis does not measure the rare-language improvements described by ElevenLabs.

Speaker diarization, which identifies who spoke, also works more reliably in larger group conversations, according to ElevenLabs. Scribe v3 Live makes that capability available in real time for the first time. For a live agent, speaker identity can become part of the operating context rather than a label added after a call.

The distinction between files and live calls carries practical weight for developers. Scribe v3 handles pre-recorded audio. Scribe v3 Live is the component meant to feed a voice agent while a caller is still on the line, according to ElevenLabs.

How does Scribe v3 compare with rival models?

Independent benchmarks from Artificial Analysis place Scribe v3 among the most accurate transcription models on the market. It edges out models from Microsoft, Mistral, xAI, Google, and OpenAI, according to Artificial Analysis. Only a preview model from Alibaba ranks higher in the cited results.

The accuracy lead is real within that field. It is not a lead over every production measure.

Artificial Analysis puts ElevenLabs mid-pack on speed and price. Models from Microsoft and xAI run many times faster at marginally higher error rates, the benchmark says. Those rivals also cost roughly half as much as ElevenLabs' models. Buyers processing large volumes of recorded audio quickly and cheaply are better off with other providers, according to the benchmark analysis.

The benchmark has a narrow boundary. Artificial Analysis measures only word errors on recorded files. It does not account for the speaker-diarization improvements, rare-language performance, or real-time behavior that ElevenLabs promotes. The results support the headline accuracy claim, but leave several product advantages outside the scorecard.

For ElevenLabs, the trade-off is straightforward. Its new model competes on precision while faster, cheaper rivals compete on throughput and cost. The company is selling that gap to developers who need fewer transcription mistakes, especially where speakers overlap or technical terms carry meaning.

The benchmark does not show how much speed or cost buyers will accept to avoid those errors.

Why does conversational context matter for live agents?

The live model tackles a problem that sits between speech recognition and conversation. Current voice agents often depend on rigid pause timers. A caller who stops briefly to think, or slowly reads a string of numbers, can be cut off mid-sentence, according to ElevenLabs.

ElevenLabs says Scribe v3 Live “analyzes conversational context alongside the audio signal” so it can tell when a speaker has actually finished. It uses the surrounding exchange as context instead of treating silence alone as proof that a turn has ended. That design targets a specific failure in live agent calls.

Artificial Analysis did not measure that behavior because its test covers recorded files only. The release also provides no figures for latency, false interruptions, or how often callers continue after a bot cuts in. Those missing measures matter because context-aware turn detection is the central promise of Scribe v3 Live.

The stakes are practical for any deployment built around voice agents. A transcript that assigns the wrong speaker creates a different product problem from an agent that ends a turn too early. ElevenLabs says both models improve speaker handling, with the live model extending it into real time for the first time.

What can developers control in the transcript?

ElevenLabs says the models can be “preloaded with thousands of specialized terms” like brand names or medications. The company says this helps stop domain-specific vocabulary from being misheard. The feature gives a deployment vocabulary chosen by its developers rather than leaving every proper noun to general recognition.

Developers can also set formatting rules for numbers and dates in natural language. A multilingual mode handles language switches in the middle of a sentence. Together, these controls let teams shape recognition output before it passes into another system, according to the feature set ElevenLabs announced.

Enterprise customers can choose isolated servers in the EU. They can also enable a zero-retention mode designed to delete data immediately after each request. For enterprise buyers, those controls add data-location and deletion choices to the model comparison. ElevenLabs is presenting them alongside transcription quality and live-agent behavior.

Artificial Analysis can measure word errors and price. It cannot measure whether a model can recognize a medication name, follow a date-formatting rule, or handle a caller switching languages mid-sentence. ElevenLabs is presenting those features as product controls, not as benchmark scores.

What does ElevenLabs' enterprise expansion add?

ElevenLabs is also expanding its enterprise business alongside the models. Under an expanded partnership with Nvidia, the company is integrating the open Nemotron 3.5 Super VL language model into ElevenAgents. It is also securing additional compute capacity running on Nvidia.

The Nvidia agreement includes work on local voice control for robots. It also includes a phone agent for Poland's public health system, according to ElevenLabs. These deployments extend the speech models beyond meeting transcription and into devices and public services.

ElevenLabs also announced an alliance with consulting firm McKinsey. The company introduced a broader partner program for system integrators. It also announced its own marketplace, where companies like Salesforce, IBM, and Zendesk offer ready-made voice agents.

The release does not detail the partnership terms or any revenue target. It does show how ElevenLabs wants to package the models: as components within a wider agent stack, supported by infrastructure, consulting partners, integrators, and a marketplace.

This expansion puts transcription quality inside a larger business. Accuracy still matters, but the company is also competing on deployment, governance, and access to ready-made agents. Artificial Analysis makes the central trade-off plain: Scribe v3 ranks near the top for accuracy, while speed and cost remain weaker than Microsoft and xAI.

ElevenLabs now has a two-model speech stack with a measurable accuracy claim: about 25 percent fewer word errors across 102 languages. It also has a live model that the company says uses conversation context to prevent mid-sentence cutoffs, plus real-time speaker diarization for the first time.

The enterprise layer adds vocabulary controls, EU isolation, zero retention, and partnerships across Nvidia, McKinsey, Salesforce, IBM, and Zendesk. Across the release, ElevenLabs is trying to make Scribe more than a transcription endpoint. The models sit inside ElevenAgents, alongside Nemotron 3.5 Super VL and a marketplace of ready-made voice agents.

The next test is whether live context and enterprise controls can offset the benchmark gap. Microsoft and xAI remain many times faster and roughly half as costly, even with marginally higher error rates in recorded-file testing.