ElevenLabs Launches Scribe v3 for Recorded and Live Speech
We cover speech to text news. ElevenLabs has launched Scribe v3 and Scribe v3 Live. The company published the news on November 10, 2026 and updated it on November 14, 2026.
Scribe v3 is built for recorded audio. Scribe v3 Live is built for live audio. Both models are available in the ElevenLabs API. Scribe v3 Live is also available in ElevenAgents.
Accuracy across many languages
ElevenLabs says Scribe v3 cuts the share of missed or changed words by 27 percent across 102 languages. The comparison is with Scribe v2. The company says the new model improved in every language it tested.
The company measured multilingual accuracy on an internal version of the FLEURS test. That test uses short, clean read speech in 102 languages. In languages the company calls easier to transcribe, German improved by 42 percent. Spanish improved by 35 percent. English improved by 23 percent. In languages the company calls harder to transcribe, Swahili improved by 34 percent. Tamil improved by 29 percent. Thai improved by 28 percent.
For live speech, ElevenLabs says Scribe v3 Live cuts the share of misses that change meaning by 43 percent. The comparison is with Scribe v2 Realtime. The share of missed or changed words falls by 31 percent.
Speaker labels and turn taking
Scribe v3 also focuses on speaker labels, which mark who said what. ElevenLabs says the model has the most accurate speaker labels of any model it tested. On recordings with three or more speakers, the company reports a 17 percent drop in missed speaker labels. The comparison is with Scribe v2.
Scribe v3 Live adds speaker labels for live transcription as well. The company says this is the first time its live model has offered that feature. The company tested the English part of a 16 language real world data set. It reports a 3.1 percent rate of missed speaker labels.
The live model also adds turn detection. It uses the context of the conversation as well as the audio. The company says this helps a voice agent wait when a caller pauses mid sentence or mid number.
Controls for terms and formats
Scribe v3 can suggest which languages a recording may contain. Users can choose whether numbers come back as digits or as words. Keyterm prompting accepts up to 1,000 terms, such as brand names, acronyms, and jargon. Transcript edits can apply format rules written in plain English, such as a standard date format.
Scribe v3 Live adds a multilingual mode for conversations that switch languages in mid sentence. It also adds a keyterm library for large vocabularies. The model checks that library when it hears a likely match.
Eneco supplies energy to more than 2 million homes and businesses in the Netherlands. It tested Scribe v3 Live on its voice agent. In real customer calls, the agent recognized Dutch postcodes correctly on the first try 40 percent more often. It recognized house numbers 15 percent more often. Ruud Huigsloot is a senior product manager at Eneco. He called the early results significant and said the work with ElevenLabs will continue.
Where teams can use the models
The models run in the United States by default. ElevenLabs also offers isolated environments in the European Union, India, and Singapore. Zero Retention Mode deletes audio and text as soon as a request completes.
Both models are covered by the company’s SOC 2, ISO 27001, and PCI DSS Level 1 programs. They are eligible for HIPAA coverage with a Business Associate Agreement. They are also available for private deployment. For health care teams, ElevenLabs is releasing a version of Scribe v3 Live tuned for medical conversations in ElevenAgents.
Existing API users can switch by changing the model ID to scribe_v3 or scribe_v3_live. We will keep tracking speech to text news as these models reach more transcription workflows.