ElevenLabs releases speech models for files and voice agents

ElevenLabs has introduced Scribe v3 and Scribe v3 Live, speech recognition models for recorded files and real-time use. They are designed for challenging audio environments, such as conversations with overlapping speech and background noise. According to Artificial Analysis, Scribe v3 is among the most accurate transcription models, but lags competitors in speed and price. Scribe v3 Live analyzes the flow of a conversation so voice agents do not interrupt someone during a brief pause. Both models are available through ElevenAPI and ElevenAgents.

ElevenLabs releases speech recognition models for files and voice agents

Accuracy and speed

Scribe v3 is designed for prerecorded audio, while Scribe v3 Live is for low-latency applications running in real time. ElevenLabs says the new Scribe v3 makes about 25% fewer word errors than the previous version and supports 102 languages. The company also notes a significant improvement in accuracy for German. Identifying who is speaking and when has become more reliable in large groups, and Scribe v3 Live makes this feature available in real time for the first time.

According to independent benchmark Artificial Analysis, Scribe v3 is more accurate than models from Microsoft, Mistral, xAI, Google and OpenAI. Only a preview version of Alibaba’s model ranked ahead of it.

Scribe v3 is the most accurate model in the Artificial Analysis test but lags somewhat in speed. Its benchmark price is $3.33 per 1,000 minutes (not shown).

Source: the-decoder.com

For those who need to process large volumes of audio quickly and cheaply, the results are less appealing: ElevenLabs sits around the middle of the rankings for speed and cost. Microsoft and xAI models run many times faster, make only slightly more errors and cost about half as much.

The benchmark counts only word errors when processing recordings. It does not evaluate speaker recognition, performance on rare languages or how Scribe v3 Live performs in real time — advantages highlighted by ElevenLabs.

How Scribe v3 Live avoids interruptions

Voice agents often rely on fixed pause timers. As a result, they may start speaking before someone has finished — for example, when the person briefly pauses to think or slowly dictates a sequence of digits. Scribe v3 Live matches the audio signal against the conversation’s context to determine whether the person has actually finished speaking.

Developers can preload thousands of specialized terms into the models — such as brand names or medications — to reduce errors in professional vocabulary. Formatting rules for numbers and dates can be specified in plain language, while the multilingual mode handles a language switch mid-sentence. Enterprise customers can use isolated servers in the EU and a no-data-retention mode, which is intended to delete data immediately after each request.

ElevenLabs’ new partnerships

Alongside the model release, ElevenLabs is expanding its enterprise business. Under an expanded partnership with Nvidia, the company is adding the open Nemotron 3.5 Super VL language model to ElevenAgents, gaining additional computing capacity on Nvidia hardware and developing local voice control for robots. ElevenLabs has also launched a phone agent for Poland’s public health system.

The company also announced a collaboration with consulting firm McKinsey, an expanded partner program for systems integrators and the launch of its own marketplace. Companies including Salesforce, IBM and Zendesk are offering ready-made voice agents there.

Dataist’s take: Scribe v3 Live targets a specific problem for voice agents — ending a user’s turn too soon. But the published benchmark does not test the model in real time, so its accuracy in this situation cannot be compared with competitors based on the data provided. The trade-off between speed, cost and accuracy also remains significant for processing large volumes of recordings.