読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · ElevenLabs重大2 分で読めるDisconfirmAI 編集部

US AI Audio Firm ElevenLabs Launches Scribe v3 and v3 Live: First Realtime Speaker Diarization, Significant Error Rate Reductions

ElevenLabs releases next-generation speech-to-text models with a 27% reduction in word errors for batch processing and the introduction of speaker diarization in realtime mode.

2026-11-10 20:00 GMT+8

US-based AI audio company ElevenLabs launched two new Speech to Text models on November 10, 2026: Scribe v3 for recorded audio and Scribe v3 Live for realtime audio. The company states that these models are designed to address challenges in real-world scenarios such as noisy environments, overlapping speakers, and mid-sentence language switching.

For batch processing, ElevenLabs reports that Scribe v3 reduces Word Error Rate (WER) by 27% compared to its predecessor, Scribe v2, across 102 languages. Improvements were observed in all tested languages, with German accuracy improving by 42%, Spanish by 35%, and English by 23%. Harder-to-transcribe languages like Swahili, Tamil, and Thai saw improvements ranging from 28% to 34%.

The realtime model, Scribe v3 Live, marks a key breakthrough. ElevenLabs claims it reduces semantic WER by 43% compared to Scribe v2 Realtime and introduces 'Speaker Diarization'—automatically labeling who said what—to realtime transcription for the first time. In recordings with three or more speakers, Scribe v3's diarization error rate is 17% lower than the previous generation. Additionally, improved 'Turn Detection' uses conversational context rather than just silence duration to determine when a user has finished speaking, preventing interruptions during mid-sentence pauses.

To meet enterprise needs, the new models offer enhanced output controls, including options to render numbers as digits or words, support for up to 1,000 keyterm prompts to improve proper noun recognition, and plain-text instructions for formatting rules. Scribe v3 Live also adds a multilingual mode supporting language switching within sentences.

Scribe v3 Live is currently available in the ElevenAgents platform, and both models can be accessed via the ElevenAPI. ElevenLabs notes that migrating from Scribe v2 requires changing only one line of code for the model ID. It is important to note that the performance data cited above is based on ElevenLabs' internal datasets and an adapted version of the FLEURS benchmark, representing vendor self-reported results that have not yet been independently verified.

ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

ソース

  • elevenlabs.ioElevenLabs↗
トピック

ElevenLabs

美国AI音频公司ElevenLabs于2026年11月10日发布新一代语音转文字模型Scribe v3及实时版Scribe v3 Live。Scribe v3在102种语言中的词错误率较前代降低27%,而Scribe v3 Live首次引入…

2 ステップ1 件更新日 2026-11-14 09:37

経緯

  1. 2026-11-10ElevenLabs正式发布Scribe v3和Scribe v3 Live模型,宣布批量处理词错率降低27%,实时模式首次支持说话人分离。elevenlabs.io ↗
  2. 2026-11-11第三方评测机构BenchLM发布分析,指出ElevenLabs的测试结果基于内部FLEURS副本及未公开的16语言数据集,目前尚无法独立复现。benchlm.ai ↗

未解決の問題

  • ElevenLabs承诺开源的16语言真实会议数据集何时正式可用?
  • 独立第三方(如Artificial Analysis)的基准测试结果是否证实Scribe v3在跨语言准确率上的领先优势?
  • Scribe v3 Live的实时说话人分离功能在高噪声或重叠说话场景下的实际表现如何?
トピック全文を読む →
リサーチを続ける →