ElevenLabs Scribe v1

Sobre ElevenLabs

Fundado em 2022 · London, UK / New York, USA

ElevenLabs was founded in 2022 by Piotr Dabkowski and Mati Staniszewski, two Polish technologists who had worked at Google and Palantir. The company became famous for high-quality multilingual text-to-speech and AI dubbing, and in February 2025 expanded into the inverse problem with Scribe v1, the company's first dedicated automatic speech recognition model. Scribe was developed in part to power ElevenLabs' Dubbing Studio (transcribe source audio, translate, then re-synthesise in the target language), and is offered as a standalone API to enterprise customers who want a single vendor for the full STT-to-TTS pipeline. ElevenLabs has raised over $280M to date, with a Series C in January 2025 at a $3.3B valuation.

Arquitetura

Proprietary encoder-decoder speech-to-text Transformer

ElevenLabs Scribe v1 is a hosted automatic speech recognition model launched in February 2025. ElevenLabs has not published a technical report, but the launch blog describes a Transformer encoder-decoder ASR architecture trained on a large multilingual speech corpus covering 99 languages, with particular emphasis on accuracy in long-tail languages where Whisper Large v3 underperforms. Scribe outperformed Whisper Large v3 and Deepgram Nova-2 in the company's published FLEURS and Common Voice evaluations across many language pairs, and ranked first overall in a head-to-head benchmark on Hindi, Mandarin, German and Italian. The model supports speaker diarisation up to 32 speakers, word-level timestamps with sub-100 ms precision, character-level confidence scores, automatic non-speech event detection ([applause], [laughter], [music]) and audio-event classification. Maximum file size is 1 GB and maximum audio length is 2 hours per request.

Parâmetros

Undisclosed

Contexto

7.200 tokens

Capacidades

• 99-language multilingual ASR including many low-resource languages
• Speaker diarisation up to 32 speakers
• Word-level timestamps with sub-100 ms precision
• Non-speech event detection ([applause], [laughter], [music])
• Character-level confidence scores
• Up to 2 hours per request, 1 GB file limit
• Direct integration with ElevenLabs Dubbing Studio (STT to translate to TTS)
• Best for: dubbing pipelines, multilingual transcription, podcast indexing, media analytics

Treinamento & licença

Not disclosed. ElevenLabs reports training on a 'large multilingual corpus' with curation for long-tail languages; data is described as a mix of licensed and crowd-sourced opt-in audio.

Licença: Proprietary commercial API. Commercial use permitted on paid plans.

Testes de segurança: Same KYC and identity-verification practices as the rest of the ElevenLabs platform; no formal red-team report for ASR.