ElevenLabs has released two speech-to-text models, Scribe v3 for pre-recorded audio and Scribe v3 Live for real-time use, and the most interesting claim is the one about interruptions. Anyone who has been cut off by a voice agent while reading out a card number knows the problem. Current systems often rely on rigid pause timers that treat a thoughtful silence as the end of a sentence. The company says Scribe v3 Live analyzes conversational context alongside the audio signal to judge whether a speaker has actually finished. If that works as described, it addresses a real flaw in how voice agents feel to use.

The accuracy claims deserve credit but also scrutiny. ElevenLabs says Scribe v3 cuts word errors by about 25 percent compared with its predecessor across 102 languages, and independent benchmarks from Artificial Analysis place it among the most accurate transcription models on the market, behind only a preview model from Alibaba. That is a strong result. Yet the same benchmark measures word errors on recorded files, so it cannot confirm the company's claims about speaker diarization in group conversations, rare languages, or the real-time behavior of Scribe v3 Live. Those improvements may be real, but they remain the company's word until someone tests them independently.

The cost picture is less flattering. The report notes that Elevenlabs lands mid-pack on speed and price, while models from Microsoft and xAI run many times faster at marginally higher error rates and cost roughly half as much. For a developer processing large volumes of audio, that gap matters more than a small edge in accuracy. Buyers should choose the tool that fits the job rather than the one at the top of a single leaderboard.

The privacy features are a meaningful step. Enterprise customers can get isolated servers in the EU and a zero-retention mode designed to delete data immediately after each request. Vendors should be expected to offer these controls as standard, and buyers should ask for written confirmation of how they work in practice.

The wider business push, including a partnership with Nvidia, a McKinsey alliance, and a marketplace listing offerings from Salesforce, IBM, and Zendesk, shows where ElevenLabs wants to compete. Enterprise voice is a crowded field, and the company's best argument will be reliability in live calls rather than headline scores. Judge the product on whether callers stop being interrupted, not on the benchmark alone.