---
title: "Scribe v3 and v3 Live launch: speech benchmarks and the batch-versus-streaming choice"
description: "Compare Scribe v3 speech-to-text benchmarks with peers, understand v3 Live's speaker-label test and choose between recorded and streaming audio workflows."
url: "https://promptszone.com/scribe-v3-live-benchmarks-speech-to-text"
published: "2026-11-12T07:00:04Z"
categories: ["New Models"]
---

# Scribe v3 and v3 Live launch: speech benchmarks and the batch-versus-streaming choice

ElevenLabs releases Scribe v3 and v3 Live. Compare matched-language transcription scores, real-time speaker labels and practical migration checks.

![Developer-reported word error rates across 102 FLEURS languages: Scribe v3 12.4%, Scribe v2 17.0%, Gemini 3.5 Transcribe 18.5%; lower is better.](https://promptszone.com/assets/uploads/313ac1bcd7b755a575a5ed03ab1ee907.png)

**Summary:**

Scribe v3 handles recordings while v3 Live handles streaming conversations. Two separate developer evaluations show why transcription and speaker-label scores need distinct comparisons.

**ElevenLabs launched Scribe v3 and Scribe v3 Live on November 10, 2026.** Both speech-to-text models are available through ElevenAPI; the Live version also powers real-time transcription in ElevenAgents. The release targets two different jobs: processing recordings and following conversations as they happen.

## Which model fits the workflow?

The documented model IDs are `scribe_v3` for recordings and `scribe_v3_live` for streaming audio. Scribe v3 supports speaker labels for up to 32 speakers, word-level timestamps, number formatting and keyterm prompting with up to 1,000 terms. Scribe v3 Live adds context-aware turn prediction, multilingual conversation handling and a keyterm library.

For a meeting archive, speaker attribution and terminology may matter more than immediate output. For a live agent, the evaluation should also include interrupted speech, mid-sentence pauses and the delay before a response begins. These are different acceptance tests, even when both products return transcripts.

## Benchmark scorecard: keep the test populations aligned

ElevenLabs’ October 23–29 internal tests compared these three recorded-audio models across the same 102 FLEURS languages. Each language was scored separately before averaging. Lower values are better; the best numeric value in each row is bold.

| Metric | Scribe v3 | Scribe v2 | Gemini 3.5 Transcribe |
| --- | --- | --- | --- |
| Word error rate | **12.4%** | 17.0% | 18.5% |
| Semantic word error rate | **4.4%** | 5.4% | 7.5% |

Word error rate counts transcription differences; semantic word error rate focuses on differences that change meaning. FLEURS uses short, clean read speech, so this comparison does not measure multi-speaker meeting performance. The launch chart’s other model comparisons use different language subsets and are not combined into this ranking.

## Real-time speaker labels use a separate test

On November 4, ElevenLabs evaluated streaming speaker attribution on English recordings from its own meetings, calls and voice-note dataset. Transcripts were aligned to a reference before scoring, isolating speaker assignment from transcription errors.

| Metric | Scribe v3 Live | Grok Voice Transcribe 2.0 | AssemblyAI Universal-3.6 Pro Realtime | Deepgram Nova-3 |
| --- | --- | --- | --- | --- |
| Aligned diarization error rate | **3.1%** | 5.9% | 6.1% | 7.3% |

This is a developer-run English evaluation, not a public multilingual benchmark. Its alignment procedure means the percentages should not be compared directly with ordinary batch diarization scores.

## What to check before switching

Use an evaluation set that reflects the intended workload: accents, overlapping speakers, product names, numbers and recording quality. Score both the transcript and downstream outcomes, such as whether an action or summary preserves the right amount, name and speaker.

For streaming applications, test turn completion separately from recognition accuracy. A transcript can be correct while the agent still responds too early. Review the current API documentation, deployment conditions and billing before migrating production traffic.

## Sources and image credit

- [ElevenLabs launch announcement and benchmark charts, November 10, 2026](https://elevenlabs.io/blog/scribe-v3). Results above are developer-reported, with test dates and populations retained.
- [ElevenLabs model documentation](https://elevenlabs.io/docs/overview/models/), checked November 12.
- Original chart based on ElevenLabs’ reported data. Official ElevenLabs provider logo from its [press kit](https://elevenlabs.io/press), following its [brand guidance](https://elevenlabs.io/brand). This is a provider-logo fallback, not a separate Scribe logo.