ElevenLabs has launched two new speech-to-text models, Scribe v3 and Scribe v3 Live, designed to handle noisy environments and real-time audio. Scribe v3 targets pre-recorded files, cutting word error rates by approximately 25% across 102 languages, while Scribe v3 Live brings real-time speaker diarization to live calls. Both models are available via the ElevenAPI developer interface and ElevenAgents platform.

According to the company, Scribe v3 Live addresses a common issue in voice agents—cutting off callers during brief pauses—by evaluating conversational context alongside audio signals to determine when a speaker has truly finished. Independent benchmarks from Artificial Analysis confirm top-tier transcription accuracy for Scribe v3, though competitors like Microsoft and xAI offer faster speeds at lower costs.

Alongside the launch, ElevenLabs expanded its enterprise presence by integrating Nvidia’s open Nemotron 3.5 Super VL model into ElevenAgents and partnering with McKinsey and major platforms like Salesforce, IBM, and Zendesk.

Why it matters

  • Scribe v3 Live improves voice agent UX by eliminating rigid pause timers and premature interruptions using context awareness.

  • Benchmarking shows a trade-off: ElevenLabs excels in accuracy and diarization, but competitors offer faster, cheaper transcription alternatives.

  • Integrations with Nvidia and enterprise platforms signal expanding deployment opportunities for custom voice agent ecosystems.

Source: the-decoder.com