Speech to Text: Transcription with Scribe

Scribe: ElevenLabs' Transcription Engine

Scribe is ElevenLabs' speech-to-text product. Scribe v2 (released early 2026) transcribes audio in 90+ languages with word-level timestamps. It is not just a transcription tool — it also tags audio events, identifies speakers, and can be told to watch for specific terms so it handles proper nouns and technical vocabulary accurately.

Scribe sits on the output side of the voice AI pipeline. You have already seen how ElevenLabs helps you create voice content. Scribe closes the loop by converting spoken content back to text — transcribing meetings, interviews, podcasts, or any audio you need in written form.

Core Features of Scribe v2

Word-level timestamps — Every word in the transcript is tagged with its start and end time in the audio. This makes Scribe's output directly usable for subtitle generation, where each line must be precisely timed, and for aligning transcripts to video editing timecodes.

Speaker diarisation — Scribe identifies and labels different speakers throughout the audio. You get a transcript that distinguishes Speaker 1 from Speaker 2 from Speaker 3. This is critical for interview transcripts, meeting notes, and any multi-speaker recording where knowing who said what matters.

Audio event tagging — Scribe detects non-speech audio events and includes them in the transcript output: [applause], [laughter], [music], [background noise]. This context is useful when the transcript will be reviewed or when you need to understand what was happening in an audio segment beyond the words spoken.

Keyterm prompting — You can provide a list of up to 1,000 terms that Scribe should watch for. Brand names, technical terms, proper nouns, and acronyms that a general model might transcribe incorrectly are handled accurately when you provide them as hints.

Scribe v2 Realtime

Scribe v2 Realtime is a separate variant built for live transcription. It processes audio as it is captured, producing transcription with under 150 ms latency. Use cases include live captioning for events, real-time note-taking during meetings, and integration with voice agents that need to convert user speech to text instantly.

The batch variant (standard Scribe v2) is better for recorded content where you want maximum accuracy and the full feature set including diarisation and keyterms. The realtime variant prioritises speed over completeness.

Common Scribe Workflows

• Podcast transcription — Upload the audio, enable diarisation, download the transcript as text or SRT. Use word-level timestamps to generate captions in your video editor.
• Meeting notes — Record a meeting, upload to Scribe, get a diarised transcript. Feed the output to a language model to generate a summary and action items.
• Subtitle file generation — Use the timestamped output to produce an SRT or VTT file for a video. The word-level timestamps make it easy to create subtitle segments that match the actual speech rhythm.
• Content repurposing — Transcribe a video or podcast episode, then use the text as the basis for a blog post, newsletter, or social content.

Language Coverage and Accuracy

Scribe v2 covers 90+ languages. Accuracy is generally strong on high-resource languages (English, Spanish, French, German, Portuguese, Japanese). On lower-resource languages or heavily accented speech, use keyterm prompting to anchor the most critical vocabulary and review the output before final use.

Try This

Find a 5-minute audio recording — a podcast clip you downloaded, a recorded voice memo, or any speech audio you have permission to use. Upload it to ElevenLabs and run Scribe v2 on it. Enable speaker diarisation and word-level timestamps. Review the output: how accurate is the transcription? How well did it separate speakers? If there were proper nouns it got wrong, try re-running with those terms added as keyterms and compare the results.