Intel / article

Introducing Scribe V2 Realtime

Source: elevenlabs.io
Author: ElevenLabs editorial sitemap
Date: Published Oct 9, 2026

Context

ElevenLabs presents Scribe v2 Realtime as speech recognition for applications that need text while a person is still talking. Its announcement reports transcription latency below 150 milliseconds and describes automatic language detection, including switches during a conversation. Applications can also control when transcript segments are finalized. Those capabilities support live captions and voice interfaces where waiting for a complete recording would interrupt the interaction. The implementation guide explains the distinction between provisional and finalized text.

Key takeaways · AI-distilled

ElevenLabs says Scribe v2 Realtime transcribes speech in under 150 ms and reports 93.5% accuracy across 30 commonly used European and Asian languages, with claimed outperformance on 500 hard samples with background noise.
Features aimed at agents include next-word and punctuation prediction, automatic language detection with mid-conversation switching, and text conditioning that continues transcription from the previous batch after a reconnect.
Developers can commit segments manually or via voice activity detection; audio input supports 48kHz PCM and μ-law, and the model is offered with EU and India data residency and a zero retention mode.

Read the source elevenlabs.io