ElevenLabs has released Scribe v3 and Scribe v3 Live, speech-to-text models for teams that transcribe meetings, support calls, clinical notes, and live voice agents. Scribe v3 is for recorded audio. Scribe v3 Live is the low latency model for realtime audio, with a median latency of 150 milliseconds, and it is already available inside ElevenAgents. The models target the audio those workflows actually get: overlapping speakers, background noise, pauses that are not the end of a sentence, and conversations that switch languages. On the FLEURS benchmark, Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, and 17% fewer diarization errors on recordings with three or more speakers. Developers can return numbers as digits or words, pass secondary language hints, bias recognition with up to 1,000 key terms, and apply formatting rules written in plain English. Scribe v3 Live adds speaker labels to realtime transcripts for the first time. ElevenLabs reports a 43% lower semantic word error rate than Scribe v2 Realtime, and turn detection that uses the full conversation rather than silence alone, so a caller pausing to read a number is not cut off. Both models are available in the ElevenLabs app and through ElevenAPI, where existing integrations switch by changing the model ID. They run in the US by default, with isolated environments in the EU, India, and Singapore. ElevenLabs, founded in 2022 by Mati Staniszewski and Piotr Dąbkowski, is known for text-to-speech and voice agents. In a pre-launch test on ElevenAgents, Dutch energy company Eneco saw postal codes recognized correctly on the first try 40% more often.