ElevenLabs' Scribe v3 Live Adds Real-Time Speaker Labels and Stops Cutting Off Callers Who Pause Mid-Number

ElevenLabs has released Scribe v3 and Scribe v3 Live, two new speech-to-text models aimed at the messy audio that trips up transcription: overlapping speakers, mid-sentence pauses and callers who switch languages. The company announced both models on November 10, 2026. Scribe v3 Live is available in ElevenAgents, and both are available in ElevenAPI under the model IDs scribe_v3 and scribe_v3_live.

The two models split along a familiar line. Scribe v3 handles recorded audio. ElevenLabs says it makes 27% fewer word errors than Scribe v2 across 102 languages. Scribe v3 Live is the low-latency version for streaming audio, and the company says it makes 43% fewer errors than Scribe v2 Realtime, measured by semantic word error rate. It is also the first Scribe realtime model with speaker diarization, meaning it labels who said what.

Five requests from the teams already using Scribe

ElevenLabs frames the release around feedback from customers who, it says, have run “hundreds of millions of hours” of transcription through Scribe. Meeting-notes teams wanted better speaker labels in fast exchanges and with three or more people. Voice-agent builders complained that silence timers cut callers off when they paused or slowed down to read out a number. Teams building live products said they were trading accuracy for speed, because realtime ran noticeably worse than batch. Customers in India wanted Hindi to match English, including when callers switch between the two mid-sentence. And many wanted more control over formatting and vocabulary.

Turn detection that reads the conversation

The most practical change for voice agents is how Scribe v3 Live decides that someone has finished speaking. Traditional systems wait for a set period of silence. ElevenLabs points out the trade-off: a short threshold interrupts people who pause mid-sentence, while a long one makes the agent feel slow. Scribe v3 Live uses the context of the conversation as well as the audio, so a caller who pauses halfway through a number is not cut off. ElevenLabs says this is the same turn detection ElevenAgents already uses, now exposed through ElevenAPI with a turn_prediction commit strategy.

Dutch energy supplier Eneco, which serves more than 2 million homes and businesses, tested Scribe v3 Live on its ElevenAgents voice agent before launch. On real customer calls, ElevenLabs says the agent recognized Dutch postcodes correctly on the first try 40% more often, and house numbers 15% more often.

Speaker labels in batch and live audio

For recorded audio, ElevenLabs says Scribe v3 has the lowest diarization error of any model it tested: 12% fewer errors than Scribe v2 overall and 17% fewer on recordings with three or more speakers. On the English portion of an internal 16-language dataset of real meetings, agent calls and voice notes, Scribe v3 Live recorded a diarization error rate of 3.1%, compared with 5.9% to 7.3% for the other realtime models tested. ElevenLabs notes that it aligned each transcript to the reference before scoring the realtime results, so those figures cannot be compared with its batch numbers. The company says it plans to open-source the 16-language dataset.

Every language improved, Greek most in realtime

Multilingual accuracy was tested on an internal version of the FLEURS benchmark, which uses short, clean read speech in 102 languages. ElevenLabs says Scribe v3 improved in every language tested: German by 42%, Spanish by 35% and English by 23%, with Swahili, Tamil and Thai improving by 34%, 29% and 28%. For Scribe v3 Live, the largest realtime gains were in Greek (77%), Danish (54%) and Croatian (45%), with Hindi improving by 36%.

ElevenLabs scored each language separately before averaging so that all 102 count equally, and cautions that results pooled across clips “aren’t directly comparable.” All benchmark figures are the company’s own.

The blog also puts the numbers in practical terms. 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.

1,000 keyterms and plain-English formatting rules

Scribe v3 can infer which languages a recording might contain and lets developers choose whether numbers come back as digits or words. Keyterm prompting accepts up to 1,000 terms such as brand names and jargon, and transcript edits apply formatting rules written in plain English, such as writing every date as YYYY-MM-DD. Scribe v3 Live adds a multilingual mode for mid-sentence language switching and a keyterm library for vocabularies too large for a list.

Both models run in the US by default, with isolated EU, India and Singapore environments for enterprise data residency. ElevenLabs is also releasing a version of Scribe v3 Live tuned for medical conversations in ElevenAgents.