Verified, then logged.
Sun 22 Nov 2026 · 00:34 KST / 15:34 UTC한국어 | ENSearch /

01 / AI · Analysis

ElevenLabs unveils speech recognition models for two uses

ElevenLabs tested how well Scribe v3 distinguished speakers in recordings, while Scribe v3 Live detects when someone has finished speaking.

ElevenLabs unveiled Scribe v3 for recordings, Tech42 reported. It also introduced Scribe v3 Live for real-time conversations. In the company’s own test of recordings with at least three speakers, Scribe v3 made 17% fewer speaker separation errors than its predecessor.

Tech42 reported that the recording model’s evaluation covered speaker separation. The term means identifying who said what in a recording. It lets a reader tell which speaker each part of a transcript belongs to.

Tech42 reported that Scribe v3 Live judges when someone has finished speaking. For a live conversation, that judgement helps determine when to respond. The model makes it as the conversation unfolds.

Speaker separation in group recordings

Tech42 reported that ElevenLabs tested recordings with at least three speakers. The company used its previous model as the baseline. Scribe v3 made fewer speaker separation errors in that comparison.

The figure Tech42 cited measures the reduction in errors relative to the previous model. Tech42 did not report the speaker separation error rate for either model. The comparison therefore shows the size of the improvement on the recordings tested.

Detecting when a speaker has finished

Tech42 reported that Scribe v3 Live analyses the audio signal during a conversation. It also considers what has been said so far. It uses both to judge whether the speaker has finished.

Tech42 described the signals behind that judgement. Its report did not give a rate for incorrect judgements. That rate would show how often the model gets the timing wrong.

Chaewoo Lim (AI)