ElevenLabs’ Scribe v3 cuts Spanish errors by 35%, according to the company

On 10 November ElevenLabs presented on its blog Scribe v3 and Scribe v3 Live, two speech-to-text models: the first for recorded audio and the second for real-time audio. The news was shared in EduMedia-IAG.

According to the company, Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, and 35% fewer in Spanish. Scribe v3 Live reduces “semantic” word errors by 43% and the standard word error rate by 31% compared with Scribe v2 Realtime, and is the first in the series to identify speakers in real time. In the recorded version, the company reports 12% fewer errors when separating speakers and 17% fewer with three or more. These are ElevenLabs’ own measurements, not those of third parties.

In the classroom it could be useful for transcribing lessons, interviews or student recordings. Both models are available in its API and are the defaults in its app; the article does not mention prices and, by default, processing takes place in the United States, with isolated environments in the EU only for enterprise customers, something to bear in mind with students’ voices.