ElevenLabs: Scribe v2

ElevenLabs Scribe v2 is the batch speech-recognition model for recorded audio, transcription, subtitling and captioning. It builds on Scribe v1 with improved handling of long recordings, pauses, changes in tone and extended silence. The result is written speech with timing information and optional speaker and audio-event annotations. Scribe v2 Realtime is a separate live-transcription edition; this model’s purpose is processing complete recordings.

The maker documents recognition across more than 90 languages and automatic language detection, including recordings that switch languages. A known language hint can help the model interpret the audio. Prepare a clear recording, retain the original language and names, and avoid severe clipping or unnecessary processing that removes speech detail. Common documented inputs include WAV, MP3, FLAC, M4A, AAC, OGG, Opus and WebM. Choose an input representation supported by the transcription implementation and preserve the source recording for review.

The standard response contains the recognized text, detected language information and a sequence of timed words. Word records can include start and end times, speaker identifiers and a recognition log-probability. These make the result useful for subtitles, searchable interviews or an editor that jumps back to the relevant sound. Keep the timed result with the source audio. For a podcast, for example, the deliverable can be an editable transcript and a word-aligned record that supports locating corrections before arranging captions.

Speaker diarization labels who speaks when in multi-speaker audio, with up to 32 speakers documented by the maker. Enable it when speaker turns are important, and supply a reasonable speaker count if that information is known. Distinguish a detected speaker label from a verified person’s name. For an interview or meeting, review turn changes against the recording, especially where voices overlap or sound similar. A label helps organize the transcript; it does not independently establish the speaker’s identity.

A recording with one speaker per channel has a different structure from several voices mixed into one channel. The producer interface can process up to five channels independently and return separate transcripts or a combined transcript ordered by time, with channel indices. Preserve that channel layout when it is meaningful. Flattening a separated call recording before transcription can discard a useful source of speaker information. Keep channel-based attribution distinct from diarization inferred from mixed audio. The combined-channel response requires timestamps and cannot include entity detection or redaction.

Keyterm prompting guides recognition of names, product terms and specialist vocabulary using context. The current batch documentation allows up to 1,000 terms, with each term kept under 50 characters and at most five words. Provide relevant spellings rather than a broad unrelated catalogue. For a technical interview, supply the product names and abbreviations likely to occur, then inspect them at their timestamps in the result. Keyterms bias recognition; they are not instructions to insert a term when the recording does not support it.

Choose the transcript’s purpose before applying cleanup. A verbatim record retains the spoken wording, while the documented no-verbatim option removes fillers, false starts and disfluencies for a more readable result. Dynamic audio tagging can identify non-speech events such as laughter or footsteps. A polished transcript may be useful for captions or reading, but it should not silently replace the raw record when repetitions, corrections or non-speech sounds are important to the task.

The maker also documents entity detection and experimental transcript editing. Transcript editing cannot be combined with entity detection, entity redaction or multichannel transcription in the producer interface. Detected entities provide structured positions for selected information in the transcript. Optional entity redaction masks detected text with replacement labels; it depends on entity detection, and the entity list is not returned when redaction is enabled. Redaction happens during transcription, using a generic replacement marker, a category label or a numbered category label. Automated redaction can miss sensitive content, so inspect the redacted transcript before sharing it. A natural-language edit returns an edited text alongside the original result, leaving the original timing and speaker record available. Keep those artifacts distinct: a corrected spelling or reformatted sentence may need a renewed check before it is aligned with the audio or used in a downstream extraction. These features process the transcript; they do not establish that every important detail was recognized correctly.

For a finished transcript, review names, numbers, negations and speaker changes by returning to the corresponding audio. Pay attention to quiet sections, overlap and language switches rather than judging quality only from fluent text. Retain uncertain wording for correction instead of completing an inaudible passage with a plausible sentence. The useful result is a checked text record with the timing and speaker information needed for its intended use; a transcript can then become input to a separate summary, translation or analysis step.