ElevenLabs drops Scribe v2 Realtime for ultra-fast voice agents

Anyone who has tried talking to a voice agent knows the awkward pause. You stop talking, wait a couple of seconds, and the illusion of a natural conversation completely falls apart.

ElevenLabs rolled out Scribe v2 Realtime, a low-latency speech-to-text model specifically built to fix that lag.

The new model transcribes speech in under 150 milliseconds. It supports more than 90 languages, including English, French, German, Italian, Spanish, Portuguese, Hindi, and Japanese. Across 30 widely spoken European and Asian languages, ElevenLabs reports an accuracy rate of 93.5%.

Predictive audio and agent-first controls

Shaving down latency in live audio is notoriously hard, but Scribe v2 Realtime introduces a few neat engineering tricks to stay ahead of the speaker.

Most notable is what ElevenLabs calls negative latency. Basically, the system predicts upcoming words and punctuation before the user even finishes the sentence, helping close the timing gap entirely.

It also handles real-time code-switching. Speakers can swap languages mid-sentence, and the system automatically detects the change without breaking stride.

For developers building in the wild, the model includes text conditioning. If a network connection drops and restarts, Scribe can pick up the transcript using context from the previous batch instead of starting from scratch.

There is also built-in voice activity detection, manual commit controls so engineers decide when to finalize a phrase, and support for 48kHz PCM and μ-law audio.

On a benchmark of 500 challenging samples packed with heavy background noise and dense information, ElevenLabs claims the model significantly outperformed all competitors.

It is available right now via the ElevenLabs API and natively inside ElevenLabs Agents. Enterprise setups also get HIPAA, SOC 2, and zero-retention modes for sensitive data.

Why it matters

If you build customer service bots, live meeting assistants, or automated phone workflows, this update is worth testing immediately.

Until now, most teams stitched together a patchy pipeline: one vendor for transcription, another for the LLM, and ElevenLabs at the tail end for voice generation. That daisy chain inevitably created compounding latency.

By pushing transcription speed below 150 milliseconds, ElevenLabs is effectively offering a one-stop shop for building ultra-responsive conversational bots.

We think this is a smart defensive move against pure speech-to-speech models. While omni models promise native voice in the future, enterprise workflows still rely on clean text transcripts for compliance, analytics, and guardrails.

The only real risk here is how fast full-duplex end-to-end models evolve. But for any team shipping production voice agents today, Scribe v2 Realtime looks like a massive upgrade for real-world responsiveness.