01 / AI · Analysis
ElevenLabs adds speaker labels to live transcription
The real-time speaker-labelling scores correct for transcription errors and cannot be compared directly with the recorded-audio results.

ElevenLabs launched transcription models with company-reported improvements across 102 languages for developers using its ElevenAPI service. Scribe v3 handles recorded audio, while Scribe v3 Live handles real-time audio and adds speaker labelling, which identifies who said what.
The announcement separates transcription accuracy from speaker-labelling accuracy. That distinction matters when reading the results: the real-time speaker-labelling evaluation uses a different scoring method from the recorded-audio evaluation.
Both models are available through ElevenAPI, the company's developer service. Understanding the reported results requires looking at how the company scores words and speakers separately, alongside the live model's handling of pauses and language changes.
English realtime diarization error rate
| English aligned transcripts | Scribe v3 Live | 3.1 | |
|---|---|---|---|
| Other realtime models lower bound | 5.9 |
Adjusted scores for English speaker labelling
ElevenLabs' real-time speaker-labelling results are company-published figures from an English-language evaluation. Its announcement describes aligning each model's transcript with a reference transcript before scoring the speaker assigned to each word. The company explained that this adjustment separates transcription mistakes, such as missing words, from speaker-assignment errors.
According to ElevenLabs, the adjustment lowered error rates for every model in the comparison. The resulting scores therefore cannot be compared directly with its recorded-audio speaker-labelling results.
The announcement also specifies that multilingual scores are calculated separately for each language and then averaged with equal weight. For real-time transcription, ElevenLabs emphasised semantic word error rate, which counts errors when the meaning changes. Standard word error rate also counts differences in how the same content is written.
ElevenLabs also announced plans to release its own evaluation dataset of meetings and calls. That planned release is separate from the company-published results in the announcement.
Conversation context for detecting the end of speech
ElevenLabs said Scribe v3 Live uses both audio and conversational context to decide when a speaker has finished. It described the feature as designed to avoid cutting off someone who pauses while reading out a number.
The announcement also describes a new mode for the live model that handles language changes within a sentence. For recorded audio, Scribe v3 offers a choice between writing numbers as digits or words.
ElevenLabs reported improvements in every language it evaluated. For Korean, however, the announcement does not disclose an individual error rate.