ElevenLabs Scribe v3: pricing, accuracy and how to use it

ElevenLabs Scribe v3 is a speech-to-text model for recorded audio, launched November 10, 2026, alongside Scribe v3 Live for real-time transcription. Scribe v3 costs $0.20 per hour of audio through the API and Live costs $0.22, and both are already the default in the ElevenLabs app, so you can try one by uploading a recording.

What ElevenLabs posted on X

@ElevenLabs introduced the pair as "our most accurate Speech to Text models yet." The post says they are built for messy audio: heavy background noise, pauses that are not the end of a sentence, and busy meetings where you need to know who said what.

A follow-up post gives the headline numbers. It says Scribe v3 Live labels speakers for the first time and "makes 43% fewer errors than its predecessor." It also says turn detection now reads the whole conversation instead of waiting for silence.

What changed from Scribe v2

These figures come from ElevenLabs' announcement and its docs. They are the company's own tests, some on a private dataset, so treat them as claims until outside benchmarks show up.

On recordings with three or more speakers, ElevenLabs says speaker labeling is 17% better than v2. The batch model also gives word-level timestamps, can write numbers as digits or words, and exports to SRT, DOCX or PDF. Files can be up to 5 GB.

The numbers are not consistent everywhere. The blog says Live's 43% figure covers 102 languages, and the docs say 44% across 73. The blog counts 102 languages overall and the docs say "90+". None of that changes whether you should try it, but it does mean you should check a figure before you quote it to anyone.

What it costs in practice

At $0.20 an hour, a 45-minute sales call costs about 15 cents to transcribe. A shop recording 100 hours of calls a month would pay about $20, or $25 with keyterms switched on. Silence is billed like speech, so a recording with long holds costs the same as one full of talking.

Transcript editing, where you give plain-English rules for cleaning up the text, costs $0.07 an hour more on recorded files. ElevenLabs calls it experimental.

Speech to Text is on the free plan's feature list. The pricing FAQ puts it at about 330 credits a minute, and the free plan comes with 10,000 credits. By our math that is roughly half an hour of audio, which is plenty for a test.

How to try it

1. Sign in at elevenlabs.io and open Speech to Text in the app. Scribe v3 is the default, so there is no model to pick.
2. Upload one of your worst recordings: the call with the truck running in the background, or the meeting where three people talk over each other.
3. Turn on speaker labels and check two things: the names it got wrong and the numbers. Addresses, part numbers and dollar amounts are where a transcript does damage.
4. If it misses your product names or street names, add them as keyterms and run the file again.

Developers who already use Scribe only need to change the model ID from scribe_v2 to scribe_v3, or from scribe_v2_realtime to scribe_v3_live. One catch from the docs: the official SDKs do not yet support turn prediction, multilingual mode or the keyterm library on Live. For those you have to call the WebSocket API directly.

Scribe v3 Live is also available inside ElevenAgents, the company's voice agent product. We covered its outbound calling tool in our ElevenAgents Campaign Manager post.

Our read

What caught our eye is the price more than the accuracy. At twenty cents an hour a business can afford to transcribe every call it takes, and then the work is deciding what to do with all that text. Speaker labels on live audio sound like a small feature, but a phone agent that can tell the customer from the person talking behind them should make fewer strange mistakes.

We would hold the accuracy claims loosely. Every vendor in this market says it leads, as we saw with Grok Voice Transcribe 2 and Nemotron 3.5 Transcribe. Run your own audio through it, with your accents and your background noise, before you believe anyone's chart. The free plan covers a test like that.

One more caution: a transcript of a customer call is customer data. ElevenLabs offers a zero-retention mode, and HIPAA use requires a signed agreement through its sales team. Sort that out before you record patients or clients. Check your state's call recording consent rules too.

What this means for your business

At this price you can keep a searchable record of every customer call and site walkthrough, and hand it to whatever does your follow-up. It only pays off if someone decides where the text goes next. If you want help working that out, New Face Design offers a free process audit. We map how calls and jobs move through your business before we recommend any tool.