ElevenLabs Scribe v3 Live: Voice Agents That Stop Cutting In

A voice agent that waits for a fixed stretch of silence will cut off a caller who pauses halfway through reading out a postcode. ElevenLabs says Scribe v3 Live fixes that by reading the conversation as well as the audio, and one utility’s test shows what that is worth. For teams building contact-center and voice-agent stacks, the catch is in the independent benchmarks: Scribe v3 ranks near the top on accuracy, but rivals run many times faster at about half the price.

What Happened

ElevenLabs released two speech-to-text models. Scribe v3, for recorded audio, makes 27% fewer word errors than Scribe v2 across 102 languages on an internal version of the FLEURS benchmark. Scribe v3 Live, the low-latency variant, cuts semantic word error rate by 43% against Scribe v2 Realtime and adds speaker diarization, labeling who said what, to real-time transcription for the first time.

Both are available in the ElevenAPI developer interface, where existing integrations upgrade by changing the model ID to scribe_v3 or scribe_v3_live, and Scribe v3 Live is available in the ElevenAgents platform. The Decoder’s report adds that ElevenLabs is expanding enterprise partnerships alongside the launch, including with Nvidia and McKinsey.

Why It Matters

Turn detection is the headline for agent builders. ElevenLabs describes the trade-off in silence-based systems: a short threshold interrupts people who pause mid-sentence, while a long one makes an agent feel slow. Scribe v3 Live uses the context of the conversation alongside the audio, so “a caller who pauses halfway through a number isn’t cut off.” ElevenLabs says this is the same turn detection ElevenAgents already used, now exposed through ElevenAPI.

The clearest evidence comes from Eneco, a Dutch energy supplier that tested Scribe v3 Live on its voice agent before launch. In real customer calls, the agent recognized Dutch postcodes correctly on the first try 40% more often, and house numbers 15% more often. On ElevenLabs’ own real-world dataset, Scribe v3 Live’s English diarization error rate was 3.1%, against 5.9% to 7.3% for the other real-time models it tested.

Vocabulary control also matters for agents that must hear account details correctly. Keyterm prompting accepts up to 1,000 terms, Scribe v3 Live adds a keyterm library for larger vocabularies, and formatting rules can be written in plain English, such as writing every date as YYYY-MM-DD.

What This Means for Builders

The independent view is more mixed. Artificial Analysis benchmarks, as summarized by The Decoder, place Scribe v3 among the most accurate transcription models, ahead of Microsoft, Mistral, xAI, Google and OpenAI and trailing only a preview model from Alibaba. But Scribe v3 lands mid-pack on speed and price: models from Microsoft and xAI run many times faster at marginally higher error rates and cost roughly half as much. The Decoder also notes the benchmark measures only word errors on recorded files, so it doesn’t capture diarization or Scribe v3 Live’s real-time behavior.

ElevenLabs frames why small gaps matter: a one-hour meeting at 150 words per minute is about 9,000 words, so the gap between a 4% and a 2% word error rate is 180 wrong words. For regulated buyers, the models run in the US by default with isolated EU, India and Singapore environments, a Zero Retention Mode, and HIPAA eligibility with a Business Associate Agreement. Stack Archive’s comparison of GPT Transcribe against ElevenLabs, Google and Mistral covers the earlier round of this accuracy race.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.