ElevenLabs has released Scribe v3 and Scribe v3 Live, speech-to-text models targeting messy audio conditions such as crosstalk and background noise across 102 languages. Scribe v3 handles pre-recorded files, while Scribe v3 Live addresses real-time, low-latency use cases; both are accessible via the ElevenAPI and ElevenAgents platform. The company claims a roughly 25 percent reduction in word errors versus its predecessor, with significant gains in German accuracy and more reliable speaker diarization in large groups, the latter now available in real time for the first time. Independent benchmarks from Artificial Analysis rank Scribe v3 among the most accurate transcription models, trailing only an Alibaba preview model and edging out Microsoft, Mistral, xAI, Google, and OpenAI. However, it ranks mid-pack on speed and price; Microsoft and xAI models run many times faster at marginally higher error rates and approximately half the cost. The benchmark does not evaluate diarization, rare-language performance, or Scribe v3 Live’s real-time behavior. Scribe v3 Live uses conversational context alongside audio signals to detect true turn endings, preventing voice agents from cutting off callers during brief pauses. Developers can preload thousands of domain-specific terms, set natural-language formatting rules for numbers and dates, and enable multilingual mid-sentence switching. Enterprise features include isolated EU servers and a zero-retention mode. ElevenLabs also expanded partnerships with Nvidia—integrating Nemotron 3.5 Super VL into ElevenAgents, securing compute, developing robotic voice control, and deploying a Polish public health phone agent—and McKinsey, while launching a partner program and a marketplace with ready-made agents from Salesforce, IBM, and Zendesk.