Dates, sentiment, formatting: Scribe v2 edits its transcripts after listening.

ElevenLabs is adding Transcript Editing to Scribe v2 and Scribe v2 Realtime. A natural-language instruction can automatically reformat, annotate, or modify a transcript while keeping the original text available separately.

One instruction after transcription

Converting every date to ISO format, expanding abbreviations on first use, or adding sentiment to every sentence can now be requested directly as part of a transcription workflow.

With Transcript Editing, Scribe v2 receives a natural-language instruction alongside the audio. The transcription is produced first, then the instruction is applied to the resulting text.

Both versions are returned separately. The `text` field retains the original transcript, while `editedtranscript` contains the modified version. Word-level timestamps also remain attached to the original text. Up to 2,000 characters of instructions

A single instruction can contain up to 2,000 characters and combine multiple transformations.

ElevenLabs lists examples including converting times to 24-hour format, expanding abbreviations such as ETA or ASAP, removing selected terms, labeling every sentence as positive, negative, or neutral, and reformatting the entire transcript as a list.

Instructions can be written in different languages, although ElevenLabs says English instructions currently work best. The edited transcript remains in the language of the original audio.

Some operations still have dedicated parameters. ElevenLabs recommends keyterm prompting for improving recognition of specific names and terms, `noverbatim` for removing filler words and disfluencies, and `numbersformat` for controlling how numbers are written. The same workflow extends to Scribe v2 Realtime

Scribe v2 Realtime supports the same type of instructions. The instruction is passed once when the connection opens.

Each committed transcript is then followed by an `editedtranscript` event. Partial transcripts are never edited, avoiding repeated transformations while speech recognition is still in progress.

Editing takes place after transcription is complete. The documentation therefore notes that the process adds latency that increases with transcript length. The original remains available

The system deliberately separates transcription from transformation. If an edit fails, Scribe retains the successful original transcript and returns a separate error for the editing step.

Structured information such as word-level timestamps, speaker labels, and additional formats also continues to describe the original transcript rather than the rewritten version.

Transcript Editing currently cannot be combined with `entitydetection`, `entityredaction`, or `usemultichannel`. A 30% surcharge on transcription

The feature is experimental and available with Scribe v2 and Scribe v2 Realtime. Using it adds 30% to the base transcription cost, with a minimum of ten seconds of audio billed per request.

On the API side, integration uses the `transcriptedit` parameter in Python and `transcriptEdit` in TypeScript. The modified text is then returned through `editedtranscript` or `editedTranscript`, depending on the SDK.