This article contains affiliate or referral links. If you sign up or buy through them, Synthetic Labs may earn a commission or referral reward at no extra cost to you.

Scribe v3 is designed for recorded audio; Scribe v3 Live is designed for real-time transcription. For content teams, the choice starts with the output: an interview transcript, timed captions and a live voice interface need different checks.

ElevenLabs announced Scribe v3 and Scribe v3 Live on 10 November. The company positions v3 for recorded audio and v3 Live for real-time use, with improvements to transcription, speaker labelling and turn detection. Its accuracy comparisons are vendor-reported results, rather than measurements from Synthetic Labs.

Scribe v3 or v3 Live?

ModelAnnounced roleUseful evaluation focus
Scribe v3Recorded audioNames, numbers, speaker attribution and editor correction time.
Scribe v3 LiveReal-time audioTurn-taking, pauses, latency and live speaker attribution.

ElevenLabs reports 27% fewer word errors for v3 than v2 across its stated 102-language evaluation, and 43% fewer errors for v3 Live than v2 Realtime. These are relative error reductions in the vendor’s evaluation, not an assertion that either model is 27 or 43 percentage points more accurate on your recordings. See the launch post and methodology discussion.

Build a small reference recording

Before changing a content pipeline, choose an interview excerpt that resembles your actual work. Include a speaker change, a technical term, a number and a natural pause. Produce a reference transcript by listening to the audio and checking the wording.

Use the same recording to compare your existing system with a candidate. Keep the audio unchanged, record the settings and retain both outputs. We have not performed that comparison for Scribe v3; this is a suggested evaluation method.

Build a correction log your editor can use

For every meaningful error, record the timecode, reference wording, model output and correction category. Keep speaker errors separate from word errors. If you use a correction-time measure, start the clock when the editor begins checking and stop when the transcript is ready for its intended use.

Download the transcript correction log. Add one row per error and retain a copy of the uncorrected output. That lets you revisit a model decision without relying on memory.

Count the errors that affect the finished piece

Mark wrong names, omitted words, incorrect numbers and speech assigned to the wrong person. Record how long it takes an editor to turn each output into an approved transcript. A small wording error and a misattributed quotation may require very different amounts of repair.

If the output will become captions, inspect the timing as well as the text. Play the finished captions against the recording and check whether each phrase appears when it is spoken. A readable paragraph does not establish usable subtitle timing.

What word error rate leaves out

Word error rate usually compares substitutions, deletions and insertions with a reference transcript. It does not tell you whether a wrong number changes the meaning of a quote or whether the right words have been assigned to the wrong speaker. Caption timing is another separate check.

If a transcription service will feed an edit or a client deliverable, use those production requirements alongside text accuracy. You can explore ElevenLabs speech-to-text and check current access and pricing before choosing a plan. Our recommendation is to trial representative audio first.

Keep live and recorded evaluations separate

A recorded interview and a live voice interface have different acceptance criteria. For a live system, a pause in the middle of a number is a useful test: does the interaction let the speaker finish? For recorded material, the priority may be accurate speaker attribution and a clean export for the editor.

Keep a correction log from real projects. Over time, it becomes a more useful buying guide than a single headline score, because it shows which errors cost your team time and which changes actually reduce that work.

More from Synthetic Labs

photo-and-voice avatar workflow