ElevenLabs launches speech-to-text that makes 27% fewer word errors

ElevenLabs has launched two new speech-to-text models, Scribe v3 for recorded audio and Scribe v3 Live for realtime audio. The company says Scribe v3 makes 27% fewer word errors than its previous model, Scribe v2. Scribe v3 Live is available today in ElevenAgents, the company's voice agent product. Both models are also available through its developer tool, ElevenAPI.

What changed

The 27% figure comes from a test across 102 languages, and ElevenLabs says Scribe v3 improved in every one. German improved by 42%, Spanish by 35% and English by 23%. Harder languages gained too: Swahili by 34%, Tamil by 29% and Thai by 28%.

Scribe v3 also gets better at telling speakers apart, a task called diarization. That means labeling who said what. ElevenLabs says it makes 12% fewer of these errors overall, and 17% fewer in recordings with three or more people.

What the live model adds

Scribe v3 Live is the realtime version, used while a conversation is happening. ElevenLabs says it is the first to tell speakers apart in real time. On the English part of its test, the error rate was 3.1%, against 5.9% to 7.3% for other realtime models it tested.

• Fewer cut-offs: it uses the context of the conversation, not just silence, to judge when someone is done talking.
• Language switching: a multilingual mode handles callers who change languages mid-sentence.
• Accuracy: ElevenLabs says it cuts meaning-based errors by 43% compared with Scribe v2 Realtime.
• Hindi: it improves by 36% over the older realtime model.

More control over the output

Scribe v3 lets developers supply up to 1,000 key terms, such as brand names, acronyms or drug names, so tricky words come out right. It can return numbers as digits or words. Formatting rules can be written in plain English, such as "write every date as YYYY-MM-DD." The live model also has a keyterm library for thousands of terms that might come up.

Who it affects

Most people will meet these models inside other products, such as meeting-note apps, call centers, clinical documentation tools and voice agents. ElevenLabs says teams building those products asked for these improvements. For developers, switching from Scribe v2 is a one-line change to the model ID.

What to keep in mind

The accuracy numbers come from ElevenLabs' own testing, including a dataset of real meetings, agent calls and voice notes in 16 languages. The company says it will open source that dataset soon. The material does not give pricing. It also does not say which apps will adopt the new models or when.