02 / IT · Analysis
ElevenLabs releases Scribe v3 transcription models
Neural Brew reported a company-published reduction in transcription errors and support for real-time speaker identification.

ElevenLabs released Scribe v3 through an API for recorded-audio transcription, with a company-published evaluation showing 27% fewer word errors than its previous model, according to Neural Brew.
Neural Brew reported two changes: fewer transcription errors and support for distinguishing speakers in real time. These address separate tasks: turning speech into text accurately and identifying who is speaking.
The report attributed the error reduction to an internal company evaluation. Its comparison scope matters when interpreting the accuracy claim alongside the live model's speaker-identification feature.
Error reduction across all evaluated languages
Neural Brew reported that ElevenLabs published results from its internal FLEURS evaluation. The improvement is a company-published benchmark figure, rather than an independent finding.
The company-published comparison covers all 102 languages in the evaluation, according to the report. That overall scope matters: the reported reduction does not establish the improvement for any individual language.
Speaker identification in live transcription
Scribe v3 Live is a model for real-time audio streaming, according to Neural Brew. It is available through an API, an interface that lets other software connect to the model.
Neural Brew reported that the live model supports speaker diarisation, which distinguishes who is speaking. That capability and the recorded-audio model's error reduction address different aspects of transcription and need to be assessed separately.