Scribe v3 Live cuts realtime transcription errors 43% and adds speaker labels

ElevenLabs has launched two speech to text models, Scribe v3 and Scribe v3 Live. The company says the recorded-audio model, Scribe v3, cuts word errors by 27% compared with Scribe v2 over 102 languages, while the realtime Live variant trims errors by 43% against Scribe v2 Realtime.

Two models, one-line switch, enterprise hosting

Scribe v3 handles recorded audio. Scribe v3 Live is the low-latency variant for realtime use, and it brings speaker diarization (labeling who said what) to realtime transcription for the first time, along with improved turn detection. Scribe v3 Live is available in ElevenAgents, the company’s voice agent platform. Both models are available in ElevenAPI, where existing integrations switch by changing the model ID to scribe_v3 or scribe_v3_live. Both are also the default in the ElevenLabs app. The announcement gives no pricing for either model.

Output control has grown. Scribe v3 can infer which languages a recording may 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, for example writing every date as YYYY-MM-DD. Scribe v3 Live adds a multilingual mode for speakers who switch languages mid-sentence, plus a keyterm library for vocabularies too large for a list. ElevenLabs says the model consults the library only when it hears a likely match.

By default both models run in the US. Isolated EU, India and Singapore environments are offered to enterprise customers with data residency needs. A Zero Retention Mode deletes audio and text once a request completes. ElevenLabs lists SOC 2, ISO 27001 and PCI DSS Level 1 coverage, HIPAA eligibility with a Business Associate Agreement, and private deployment. A version of Scribe v3 Live tuned for medical conversations is also coming to ElevenAgents for healthcare teams.

What the benchmark numbers say

All figures below are ElevenLabs’ own, from an internal version of the FLEURS benchmark and a dataset of real meetings, agent calls and voice notes in 16 languages that the company says it will open source soon. None has been independently verified.

• Scribe v3 lowers word error rate in every one of the 102 languages tested. German improved 42%, Spanish 35%, English 23%, Swahili 34%, Tamil 29% and Thai 28%.
• On diarization, the company reports an overall drop of 12% in errors against Scribe v2, and a 17% drop for recordings with three or more speakers.
• On the English part of its 16-language dataset, Scribe v3 Live scores a diarization error rate of 3.1%, while the other realtime models tested land between 5.9% and 7.3%. The company says this score aligns transcripts to the reference first, so it cannot be compared with the batch results.
• Against Scribe v2 Realtime, Scribe v3 Live lowers semantic word error rate by 43% and word error rate by 31%. Greek improves 77%, Danish 54%, Croatian 45% and Hindi 36%.

The method matters. ElevenLabs scored each language separately before averaging, so all 102 count equally, and it warns that results pooled into a single score tend to look lower and are not directly comparable. The comparison models are not named in the text, and the batch gains rest on FLEURS, which the company describes as short, clean read speech.

A bid to own the whole voice stack

The launch reads as a platform move more than a model release. ElevenLabs already bundles speech to text, text to speech and turn taking into a flat rate for ElevenAgents: according to our earlier coverage, every paid plan works out to $0.08 per platform minute, with transcription included rather than priced separately. A more accurate in-house transcriber improves that package without adding a line item, and the turn detection is what ElevenAgents already runs, now exposed through the API.

It also sits beside a different bet. Last month ElevenLabs introduced Eleven Realtime, speech-to-speech models that replace the usual chain of recognition, language model and synthesis, with v1 Max scoring 98.8% on Big Bench Audio, a result ElevenLabs attributes to Artificial Analysis. Scribe v3 Live backs the opposite approach: pipelines that keep a transcript, which many products still need for notes, compliance records and clinical documentation. Shipping both lets the company sell to either camp.

The one customer result offered is from Eneco, a Dutch energy supplier serving more than 2 million homes and businesses. In real calls on its ElevenAgents voice agent, Dutch postcodes were recognized correctly on the first try 40% more often and house numbers 15% more often, according to ElevenLabs. Senior Product Manager Ruud Huigsloot said:

“It’s great to see significant results this early, and we are excited to continue collaborating.”

That is a single pilot, without absolute accuracy figures. Teams weighing a switch from Scribe v2 should test on their own audio, particularly noisy calls and overlapping speech, before relying on the headline percentages.