A transcript can read smoothly and still cause trouble. A promise gets assigned to the wrong colleague. A product name becomes an ordinary word. A price looks plausible until someone checks the recording. These mistakes tend to appear during the work that follows transcription, when the text is being turned into notes, a document or a customer record.

ElevenLabs released Scribe v3 and Scribe v3 Live on November 10, 2026. Both are available through ElevenAPI, and Live is also available in ElevenAgents. This article draws on public documentation. We have not tested either API.

Start with the job the transcript has to do

For a completed interview or meeting recording, the recorded-audio model is the natural starting point. A live captioning service or a voice application needs streaming transcription, with its own timing requirements.

An editor may happily wait for a transcript if it saves time during cleanup. A caller will notice when a voice service jumps in halfway through an answer. Those experiences call for different tests, even when both applications end up producing text.

The details worth checking

The model documentation lists speaker diarization and configurable number formatting for Scribe v3, alongside context-aware turn prediction for Live. The transcription documentation also describes vocabulary prompting and mixed-language features.

Speaker labels matter whenever the words carry a responsibility. If three colleagues discuss a deadline, the person agreeing to it needs to be identified correctly. An evaluation should include a way to move from a questionable sentence back to the audio. A polished summary can otherwise conceal a speaker mix-up.

Numbers deserve a separate pass. Dates, prices, reference codes and telephone numbers often carry more operational weight than the rest of a sentence. Consistent formatting can make a transcript easier to reuse in a spreadsheet, while the underlying value still needs checking.

Vocabulary is another useful test case. Bring a small list of the product names and abbreviations that occur in your own recordings. Compare the output with and without that context. A demonstration with unfamiliar material tells you much less about the cleanup your team will face.

Pricing gives you a starting point

On November 12, 2026, the official API pricing page listed Scribe v3 at $0.20 per hour and Scribe v3 Live at $0.22 per hour, excluding taxes. Check account allowances, concurrency and applicable billing conditions before planning a deployment.

For a small product, API usage is one line in the budget. Uploads, storage, retries, exports and review time also affect the cost of a usable transcript. During a pilot, record both the processing cost and the minutes spent correcting each file.

Follow the text all the way to the place it will be used. Can someone turn it into a readable document? Are speaker changes easy to follow? Is finding the relevant moment in the recording straightforward? Those small interactions have a direct effect on whether people keep using a transcription tool.

A manageable first test

Choose a few recordings you have permission to use: clear solo speech, a discussion with several speakers, and a conversation containing names or numbers. Create a short human-checked reference and keep the audio available for comparison.

Log errors in separate groups: missing words, wrong words, speaker assignments, numbers and specialist terms. Test each language you actually need. A multilingual claim gives you a reason to investigate; your recordings supply the evidence for your workflow.

For a live application, include hesitation and pauses within sentences. Watch when the system decides the speaker has finished, then check whether the resulting text remains complete.

Use de-identified samples where possible, and confirm consent and data-handling requirements before uploading recordings involving other people.

Scribe v3 offers a useful reason to revisit the tedious parts of transcription. The practical question is how much correction remains between the first transcript and the document someone can use. A small, carefully chosen pilot should make that easier to judge.