ElevenLabs has unveiled Scribe v3, a new speech-to-text model that the company says cuts word errors by 27% compared with its predecessor across 102 languages. The launch was part of a package of 13 new products and upgrades announced at the company's New York summit and reported on 12 November, as the voice AI firm moves beyond text-to-speech into the wider technology that powers spoken interaction.

Scribe v3 is built for recorded audio: meetings, interviews, podcasts and video. Alongside it comes Scribe v3 Live, a streaming variant designed for real-time use, which ElevenLabs says responds with around 150 milliseconds of latency. Together, the two models close a gap in the company's line-up. Until now ElevenLabs was best known for generating lifelike speech; Scribe v3 is about understanding it.

What Scribe v3 does differently

The headline figure is the error rate. On the company's own evaluations, Scribe v3 reduces the word error rate by an average of 27% against Scribe v2 across the 102 languages it supports. The model is also designed for conversations in which speakers switch between languages mid-sentence, moving between English, Thai, Spanish or Hindi without losing track, according to reporting on the launch.

For editors and producers, the more practical change may be speaker labelling. Scribe v3 can track up to 32 different speakers in a single recording, attributing each line to the right person even when people talk over one another. That matters in group discussions and panel recordings, where sorting out who said what has traditionally been one of the slowest parts of turning audio into text.

The model also tags sounds that are not words. Laughter, applause, sighs and whispers are identified and time-stamped, so a transcript carries some of the atmosphere of the original recording rather than just its vocabulary. Those tags can be exported into subtitle files, which video editors can use to time captions and on-screen graphics.

Built for live voice agents

Scribe v3 Live is aimed at a different job: listening while a conversation is still happening. It runs over bi-directional web connections and is designed for autonomous voice agents that need to hear a caller, understand them and respond without an awkward pause.

Its standout feature is smart turn detection. Rather than cutting a caller off at the first pause, the model analyses pitch and sentence structure to judge whether the person has actually finished speaking. That is the kind of detail that decides whether a voice agent feels helpful or infuriating, particularly when a caller is reading out a long number or thinking mid-sentence.

Scribe v3 is available through ElevenLabs' developer platform and its ElevenAgents product, and Scribe v3 Live is available in ElevenAgents now. The models arrive as voice agents move from demonstrations into everyday customer service, recruitment and operations work, where transcription errors quickly become wrong answers.

Part of a wider package

Scribe v3 is one of three voice models ElevenLabs says its researchers have shipped in the past month. The others are Eleven v4, a text-to-speech model designed to read a script with more expressive pacing, and Eleven Realtime, a speech-to-speech model for fast voice agents.

The company has also built a healthcare version, Scribe v3 Medical, tuned to recognise medical terminology, and says it plans further models for industries such as financial services. The idea is that a transcription model trained on a hospital's language will make fewer mistakes than a general one, and the same logic applies in other specialist fields.

The wider announcement included an Interaction Engine to orchestrate voice models and language models inside a conversation, routing simple steps to cheaper models and harder ones to more capable systems. ElevenAgents, the company's agent platform, gained memory, so returning customers do not have to repeat themselves, and a routines layer to run multi-step work before, during and after calls.

One caution applies to the numbers. The 27% improvement, the latency figure and the speaker counts all come from ElevenLabs' own testing, and no independent evaluation has yet been published. The direction of travel is clear enough: speech recognition is becoming a core part of the agent stack rather than an afterthought. Whether Scribe v3 sets a new standard, as its early coverage suggests, will depend on how it performs on other people's audio, not just the company's own test sets.