ElevenLabs Releases Speech-to-Text Models Scribe v3 and Scribe v3 Live: The Company Says Word Errors Fall 27% and 43% From the Previous Generation, the Realtime Model Adds Speaker Diarization, and Hourly Pricing Is Unchanged

ElevenLabs released two speech-to-text models on November 10: Scribe v3 for recorded audio and Scribe v3 Live for realtime transcription. The company says that on an internal version of FLEURS covering 102 languages, Scribe v3 makes 27% fewer word errors than Scribe v2 and Scribe v3 Live makes 43% fewer than Scribe v2 Realtime; the realtime model adds speaker diarization, turn detection that uses the whole conversation, a multilingual mode for mid-sentence language switches and a keyterm library of up to 10,000 terms. Both models are in the API as scribe_v3 and scribe_v3_live, and the pricing page lists them at $0.20 and $0.22 per hour, the same as the previous generation. All accuracy figures are the company's own tests.

Batch and realtime models move up together

ElevenLabs released two speech-to-text models on November 10: Scribe v3 handles recorded audio, and Scribe v3 Live targets realtime uses such as voice agents. Both are available in ElevenAPI under the model IDs scribe_v3 and scribe_v3_live, and ElevenLabs says moving from v2 only requires changing the ID. The official API pricing page lists Scribe v3 at $0.20 per hour and Scribe v3 Live at $0.22 per hour, the same as the previous generation.

The company's accuracy figures

The blog post says that on an internal version of FLEURS covering 102 languages, Scribe v3 makes 27% fewer word errors than Scribe v2, with German, Spanish and English down 42%, 35% and 23%; Scribe v3 Live's semantic word error rate is 43% lower than Scribe v2 Realtime's and its standard word error rate 31% lower. For diarization, Scribe v3 makes 12% fewer errors overall and 17% fewer on recordings with three or more speakers. FLEURS consists of short, clean read speech; diarization was measured on ElevenLabs' own 16-language dataset, which the company says it will open-source later. All of these are the vendor's own tests and have not yet been reproduced by third parties.

Realtime transcription learns who is speaking

Scribe v3 Live adds realtime speaker diarization, and its turn detection now uses the whole conversation to tell a mid-sentence pause from the end of a turn. A multilingual mode handles switching languages mid-sentence, and a keyterm library holds up to 10,000 proper nouns. The batch model adds language hints, an option to output numbers as digits or words, and keyterm prompting of up to 1,000 terms. ElevenLabs also says a version of Scribe v3 Live for medical conversations is in preparation. For voice support and meeting-notes apps already built on Scribe v2, comparing the new model ID on their own recordings is a safer basis for switching than the vendor's figures alone.

via: ElevenLabs Blog: Scribe v3 and Scribe v3 Live, ElevenLabs API pricing, ElevenLabs Docs: Models