ElevenLabs launches Scribe, an ASR model taking on Whisper across 99 languages

ElevenLabs has spent the last two years mastering synthetic voices. Now, the company is finally giving its platform a pair of ears.

The audio AI giant just announced Scribe, its very first speech-to-text model, claiming it takes the crown as the world's most accurate transcription tool.

Tackling 99 languages and non-speech sounds

Scribe launches with support for 99 languages right out of the box. Instead of spitting out raw blocks of text, it delivers a structured JSON payload packed with word-level timestamps, speaker diarization, and audio event tags, meaning it can detect and label ambient sounds like laughter.

According to benchmarks across FLEURS and Common Voice, Scribe consistently outperformed heavy hitters including OpenAI's Whisper Large V3, Google's Gemini 2.0 Flash, and Deepgram Nova-3.

ElevenLabs reported leading accuracy in major languages, with a 98.7% score in Italian and 96.7% in English across diverse audio types like meetings, movie subtitles, and songs.

Where it really stands out is in historically underserved languages. In Serbian, Cantonese, and Malayalam, where rival transcription engines frequently suffer word error rates above 40%, Scribe significantly cuts down mistakes.

Developers can access Scribe immediately through the new Speech to Text API, while creators can upload audio and video files straight into the ElevenLabs web dashboard. A low-latency version built for real-time applications is expected shortly.

Why it matters

This is the missing half of the equation ElevenLabs desperately needed.

Until now, anyone building conversational bots or automated voice agents with ElevenLabs had to stitch their text-to-speech together with external tools like Whisper or Deepgram. By shipping a top-tier ASR model, ElevenLabs is closing the loop, paving the way for an all-in-one conversational voice pipeline under a single roof.

If you deal with multilingual audio, video subtitling, or difficult dialects like Cantonese, Scribe looks well worth an immediate spin. Word-level sync and event markers save serious cleanup time.

That said, the battle for voice isn't just about batch accuracy; it's about speed. Until ElevenLabs ships its promised low-latency model, incumbents like Deepgram still hold the high ground for live phone agents and real-time interaction. But as a first move into transcription, Scribe is a serious shot across the bow.