ElevenLabs 发布 Scribe v3，转录与说话人分离准确率提升
ElevenLabs 的 Scribe v3 把说话人分离做到了实时流式里，错误率降了 43%，还带一万个术语的自定义词库。做会议记录、客服语音的团队可以对比下自家在用的 Whisper 或 AssemblyAI，语音 Agent 开发者尤其该看看。
ElevenLabs 推出语音转录模型 Scribe v3，官方称其转录和 diarization（标注谁在何时说话）准确率超过其测试过的所有模型。Scribe v3 Live 首次支持流式场景下的 diarization，错误率比上一代降低 43%。
技术上，v3 的轮次检测改为基于整段对话上下文判断，不再只依赖静音。新增关键词库最多可存 10,000 个术语，覆盖完整的法律或医学词表；用户还能用自然语言英文指令直接编辑转录稿。
在实际应用中，客户 Eneco 基于 ElevenAgents 搭建的语音 Agent，对荷兰邮政编码首次识别正确率提升了 40%。对做实时会议记录、客服质检和垂直行业语音应用的开发者来说，diarization 进入实时场景和术语库机制是直接可用的能力。
Scribe v3 has the best transcription and diarization (labeling who said what, when) accuracy of any model we tested. Scribe v3 Live supports diarization for the first time and makes 43% fewer errors than its predecessor. Turn detection now uses the whole conversation rather than silence alone. A new key term library holds up to 10,000 terms, enough for a full legal or medical vocabulary, and Scribe v3 supports transcript edits in plain English. In ElevenAgents, @Eneco ’s voice agent captured Dutch postal codes correctly on the first try 40% more often.📊 2 ⚡
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include ElevenLabs, Scribe v3, 语音识别, diarization.
Source attribution: curated from ElevenLabs (@elevenlabsio).
