ElevenLabs launches Scribe v3 transcription models
ElevenLabs released Scribe v3 and Scribe v3 Live, two speech-to-text models it says cut word errors by 27% and 43% from v2, and the Live model adds realtime speaker labeling.
ElevenLabs announced two new speech-to-text models. Scribe v3 handles recorded audio, and Scribe v3 Live is a low-latency variant for realtime use. Scribe v3 Live is available now in ElevenAgents. Both models are in ElevenAPI, where the company says moving from Scribe v2 means changing one line, the model ID.
On an internal version of the FLEURS benchmark, which covers 102 languages, ElevenLabs says Scribe v3 cuts word error rate (WER) by 27% compared with Scribe v2 and does better in every language it tested. German improved by 42%, English by 23% and Swahili by 34%. ElevenLabs says Scribe v3 Live makes 43% fewer errors than Scribe v2 Realtime, a figure the company reports as semantic WER, which counts an error only when the meaning changes.
Speaker labeling, known as diarization, was a main focus. ElevenLabs says Scribe v3 makes 17% fewer diarization errors than v2 on recordings with three or more speakers. Scribe v3 Live is the first version to offer diarization in realtime. The company reports a 3.1% diarization error rate on English audio, compared with 5.9% to 7.3% for other realtime models it tested. The Live model also detects when a speaker has finished talking from the conversation's context, not just from silence, and adds a mode for speakers who switch languages mid-sentence.
Other new controls include keyterm prompting for up to 1,000 terms, a larger keyterm library, and formatting rules written in plain English.
All benchmarks come from ElevenLabs. The diarization tests used an internal dataset of real meetings, calls and voice notes in 16 languages, which the company says it will open-source soon. Pricing was not covered in the material reviewed.