ElevenLabs releases Scribe v3 and Scribe v3 Live speech-to-text models

ElevenLabs has released two speech-to-text models, Scribe v3 and Scribe v3 Live, built for messy conditions such as people talking over each other in meetings or phone calls with background noise. Scribe v3 is designed for pre-recorded audio files, while Scribe v3 Live targets low-latency real-time applications. Both are available now through the ElevenAPI developer interface and the ElevenAgents platform.
According to ElevenLabs, Scribe v3 cuts word errors by about 25 percent compared with its predecessor across 102 languages, and the company says German accuracy improves significantly. Speaker diarization, the labelling of who is speaking, also works more reliably in larger group conversations, per the company, and Scribe v3 Live makes it available in real time for the first time.
Independent benchmarks from Artificial Analysis place Scribe v3 among the most accurate transcription models on the market. It edges out competitors from Microsoft, Mistral, xAI, Google and OpenAI, trailing only a preview model from Alibaba. Speed and price are a different story: ElevenLabs lands mid-pack. Models from Microsoft and xAI run many times faster at marginally higher error rates and cost roughly half as much, so anyone processing large volumes of audio quickly and cheaply is better off elsewhere. The benchmark also measures only word errors on recorded files, so it does not capture the improvements ElevenLabs touts in speaker diarization, rare languages or the real-time behavior of Scribe v3 Live.
Scribe v3 Live is aimed at a flaw in current voice agents, which often rely on rigid pause timers and cut callers off mid-sentence when they stop briefly to think or slowly read out a string of numbers. The new real-time model analyzes conversational context alongside the audio signal so it can tell when a speaker has actually finished.
Developers also get more control over the final transcript. The models can be preloaded with thousands of specialized terms, such as brand names or medications, to avoid mishearing domain vocabulary. Formatting rules for numbers or dates can be set in natural language, and a multilingual mode handles language switches mid-sentence. Enterprise customers get isolated servers in the EU and a zero-retention mode designed to delete data immediately after each request.
Alongside the models, ElevenLabs is expanding its enterprise business. Under an expanded partnership with Nvidia, it is integrating the open Nemotron 3.5 Super VL language model into ElevenAgents, securing additional compute capacity running on Nvidia, developing local voice control for robots, and running a phone agent for Poland's public health system. It also announced an alliance with consulting firm McKinsey, a broader partner program for system integrators, and its own marketplace where companies like Salesforce, IBM and Zendesk offer ready-made voice agents.
Key facts
- ElevenLabs released Scribe v3 for pre-recorded audio and Scribe v3 Live for low-latency real-time use, both available now via ElevenAPI and ElevenAgents.
- The company claims Scribe v3 cuts word errors by about 25 percent versus its predecessor across 102 languages, with significantly better German accuracy.
- Artificial Analysis ranks Scribe v3 above models from Microsoft, Mistral, xAI, Google and OpenAI, behind only an Alibaba preview model, but mid-pack on speed and price; Microsoft and xAI models are many times faster and cost roughly half as much.
- Scribe v3 Live analyzes conversational context alongside the audio signal to avoid cutting callers off during brief pauses.
- ElevenLabs also announced an expanded Nvidia partnership, a McKinsey alliance, a partner program for system integrators and a marketplace of ready-made voice agents.
Why it matters
Voice agents that interrupt callers who pause to think or read out a long number are a familiar frustration. Scribe v3 Live is ElevenLabs' answer: instead of a fixed pause timer, it uses conversational context together with the audio to judge whether a speaker has really finished. Scribe v3 Live also brings speaker diarization to real-time use for the first time, according to the company. On the recorded side, an independent benchmark puts Scribe v3 near the top for accuracy.
Who it affects
Developers building voice agents and transcription features on ElevenAPI or ElevenAgents are the direct audience, along with enterprise customers who need EU-isolated servers or zero data retention. Teams transcribing meetings, calls with background noise or multi-speaker conversations are the use cases the models are built for. Companies that need cheap, fast bulk transcription are pointed by the benchmark toward Microsoft or xAI models.
How to use it
Both models are available now through the ElevenAPI developer interface and the ElevenAgents platform. Scribe v3 suits pre-recorded files; Scribe v3 Live suits low-latency real-time applications such as phone agents. Developers can preload thousands of specialized terms like brand names or medications, set formatting rules for numbers and dates in natural language, and turn on a multilingual mode that handles language switches mid-sentence. Enterprise customers can get isolated EU servers and a zero-retention mode. No pricing figures are given.
How solid is it
The accuracy ranking comes from Artificial Analysis, an independent benchmark, which puts Scribe v3 among the most accurate models, behind only a preview model from Alibaba. The 25 percent error reduction, the German improvement and the diarization gains are ElevenLabs' own claims; the source does not say which predecessor model is meant and gives no German figures. The benchmark measures only word errors on recorded files.
Risks and caveats
The benchmark does not cover diarization, rare languages or the real-time behavior of Scribe v3 Live, so the headline feature of the live model has no independent test reported. Scribe v3 is slower and pricier than some rivals: Microsoft and xAI models run many times faster at marginally higher error rates and cost roughly half as much. No concrete word-error rates, latency figures or prices are given, and no financial terms are disclosed for the Nvidia or McKinsey partnerships.