Scribe v2 Realtime in ElevenLabs Agents

Scribe v2 Realtime is an ultra-low-latency Speech to Text model designed for real-time agents. It transcribes speech in just 30-80ms, tailored for agent environments where speed, accuracy, and conversational context awareness matter.

While typical STT systems work well in quiet test environments, their performance tends to falter in real-world settings when dealing with noise, diverse accents, and identifiers like names, emails, and IDs. ElevenLabs explains that Scribe v2 Realtime was trained to address this problem.

In internal testing, comparisons were conducted using hundreds of challenging English conversation samples. The company stated that it captured user intent more accurately even in environments with poor audio quality, varied accents, and mixed filler words.

• Multilingual performance was also emphasized.
• On the FLEURS multilingual benchmark across 30 languages, it reportedly recorded the lowest WER among low-latency ASR models.
• The focus was particularly on securing both fast responsiveness and accuracy together in global deployment environments, including Spanish, Portuguese, and Hindi.

It is already integrated into ElevenLabs Agents and can be enabled in Advanced configuration. In other words, the real-time transcription model can be used directly from the agent settings without any separate setup.