Screenshots show the official pages on the capture date and may differ from the pages at publication. Select an image to enlarge it.
On November 10, 2026, the voice AI company ElevenLabs released "Scribe v3," an AI that turns speech into text. It comes in two versions, Scribe v3 for recorded audio and "Scribe v3 Live" for real-time use, which transcribes conversations and other speech as they happen.
A big jump in accuracy
According to ElevenLabs, Scribe v3 makes 27% fewer word errors than the previous Scribe v2. This comparison covered 102 languages, and the company says accuracy improved in every language it tested. By language, errors fell by 42% in German, 35% in Spanish, 23% in English and 28% in Thai. It is also better at telling who said what in recordings with several speakers, with errors down 17% in conversations among three or more people.
The real-time Scribe v3 Live makes 43% fewer errors that change the meaning than the previous real-time version. For the first time, it can also tell speakers apart in real time. On the English portion of ElevenLabs' own evaluation data, its speaker error rate was 3.1%, lower than the other real-time models ElevenLabs tested (5.9% to 7.3%), the company says.
Main features
The main features are as follows. The first three are features of Scribe v3 for recorded audio.
- You can register up to 1,000 specialized terms and proper nouns, making them easier to transcribe correctly (the real-time version has a way to register even more terms)
- You can choose whether numbers are written as digits or as words
- You can tell it how to format the transcript in plain language, such as "format this as ..."
- The real-time version handles conversations that switch languages mid-sentence
- It judges whether a speaker has finished talking from the flow of the conversation, not just from silence
The accuracy comparisons were measured by ElevenLabs itself.
Where you can use it
Scribe v3 and Scribe v3 Live are now the default models in ElevenLabs' apps. Developers can switch simply by changing the model name in the API. Scribe v3 Live can also be used in ElevenLabs' service for building conversational AI agents.
For businesses, ElevenLabs offers settings that keep no data, processing in isolated environments in the EU, India or Singapore in addition to the default US, and contracts that comply with US medical information rules (HIPAA). Pricing was not given in the announcement, so you need to check ElevenLabs' pricing page.
CoAI's take
ElevenLabs is known for text-to-speech, but it is now also pushing high accuracy in transcription. The company says accuracy improved across 102 languages, so improvements can be expected for transcribing recordings of meetings and interviews, including in Japanese. However, the announcement does not give the size of the improvement for Japanese, so the safest approach is to test it on your own audio.
