Scribe V3 Speech-to-Text by ElevenLabs

Overview

Scribe V3 is ElevenLabs’ speech-to-text model for recorded audio and video. It converts speech into readable text, then attaches timing and structure so you can search, quote, and subtitle content. You can enable speaker labeling for multi-person recordings and tag non-speech sounds like laughter. The result works well for meeting notes, call reviews, and caption workflows where small errors cause downstream work.

What you can build

• Speaker-labeled meeting transcripts for searchable notes and action items
• Podcast and interview transcripts with non-speech event markers for editors
• Caption and subtitle prep using word-level timestamps
• Customer support call transcripts for QA review and analytics
• Lecture transcripts for study notes and accessibility
• Research transcripts for qualitative coding with time references

Inputs

• An audio or video file to transcribe. Common formats include MP3, WAV, M4A, AAC, FLAC, OGG, OPUS, and WebM, plus video formats like MP4 and MOV. Files can be up to 5 GB.
• An optional language hint as an ISO-639-1 or ISO-639-3 code (for example, en or tr). Leave it blank to auto-detect the spoken language.
• An optional toggle to label who is speaking. When enabled, the transcript includes speaker labels at speaker changes.
• An optional maximum speaker count from 1 to 32. Use this only when speaker labeling is enabled. Leave it blank to let the model infer the count.
• An optional toggle to mark non-speech sounds in the transcript, such as laughter or applause.
• An optional list of key terms you want transcribed accurately. Provide them comma-separated or one per line. You can supply up to 1,000 terms. Each term must be 50 characters or less and no more than 5 words.
• An optional cleanup mode that removes filler words, false starts, and non-speech sounds to produce a cleaner reading transcript.

Outputs

• A structured transcription result in JSON.
• The full transcript as plain text.
• A time-aligned sequence of items that reconstructs the transcript. Each item can include: - The token text (words, spacing, and detected non-speech events)
  • Start and end times in seconds when timestamps are enabled
  • A speaker label when speaker labeling is enabled
  • Per-token confidence information
• Detected language information when language detection is used.

Limitations

• Very short clips can fail. The minimum supported audio length is about 100 ms.
• Speaker labeling is capped at 32 speakers. Real recordings with crosstalk can confuse attribution.
• Speaker labels are best-effort. Overlapping speech, background noise, and distant microphones reduce accuracy.
• Key terms have strict constraints. Terms longer than 50 characters or over 5 words won’t work well.
• Non-speech event tags depend on context. Loud music or crowd noise can cause extra or missed tags.
• Low-quality inputs can degrade results. Clipping, heavy compression artifacts, and inconsistent volume often harm timestamps and word choice.

Safety & compliance

• Transcription can include sensitive personal data. Get consent and follow applicable recording and privacy laws.
• If you handle regulated health data, you may need a Business Associate Agreement (BAA) for HIPAA workflows.
• Don’t treat speaker labels as identity proof. They indicate diarized speakers, not verified individuals.
• Review transcripts before publishing. Misheard names, numbers, and medical terms can cause real harm.