ElevenLabs releases Scribe v3 and Scribe v3 Live speech-to-text models, plus expanded Nvidia and McKinsey enterprise partnerships

ElevenLabs released two speech-to-text models: Scribe v3 for recorded audio and Scribe v3 Live for real-time use, available through its ElevenAPI and ElevenAgents platform. The company claims about 25% fewer word errors than its predecessor across 102 languages. It also lists real-time speaker separation in Live, end-of-turn detection that uses conversation context rather than a fixed pause timer, custom vocabulary, mid-sentence language switching, EU-isolated servers and a zero-retention mode. The same announcement bundled an expanded Nvidia partnership, a McKinsey alliance, a partner programme and a marketplace with Salesforce, IBM and Zendesk.

Teams building phone and voice agents are the main audience. Agents that cut off callers who pause to think or read out numbers are a common complaint, so context-aware turn detection could improve call completion if it works as described. The EU hosting and zero-retention options matter for regulated buyers. The impact is incremental: it is a better component, not a change in what voice agents can do. The accuracy figure is the vendor's own, and the coverage notes that an independent benchmark (Artificial Analysis) ranks it highly but shows weaker speed and price than Microsoft and xAI. Anyone choosing a provider should test it on their own audio.

The "25% fewer errors" figure is a comparison with ElevenLabs' own earlier model, not with rivals, so it says little about its standing in the market. The Nvidia, McKinsey and marketplace announcements are partnership packaging with no measurable results yet. The headline benefit, fewer interrupted callers, comes with no published interruption rates, and the reported benchmark ranking is second-hand through The Decoder, with no primary links supplied.

1. 01Independent benchmark updates (Artificial Analysis and similar) showing word error rate, latency and price for Scribe v3 and v3 Live against Microsoft, xAI and Alibaba on noisy, multi-speaker and accented audio.
2. 02Developer or customer reports with real numbers on interruption rates, dropped turns and call completion before and after switching, rather than ElevenLabs case studies.
3. 03Whether the Nvidia, McKinsey and marketplace announcements produce named live deployments (for example the Polish public health phone agent) or published pricing within the next one to two quarters.