ElevenLabs speech to text (Scribe): transcribe calls and meetings

This page may contain affiliate links from which Blur Studio earns commissions.

Scribe is ElevenLabs' speech to text tool, part of its broader ElevenCreative content platform, that converts recorded audio, calls, meetings, interviews, into written text. It's the reverse direction of ElevenLabs' better-known text to speech product: instead of turning text into a voice, Scribe turns a voice recording into accurate, structured text. This guide covers its real accuracy claims by language, its feature set, and how it compares to a well-known free alternative, Whisper.

Accuracy and features

ElevenLabs describes its current version, Scribe v2, as offering "state of the art accuracy" across 90+ languages, though accuracy genuinely varies by language. Its own documentation breaks this down by word error rate (WER), the standard way transcription accuracy gets measured, lower is better:

• Excellent (5% WER or lower): 34 languages, including English, French, German, Spanish and Japanese.
• High accuracy (5 to 10% WER): 20 languages, including Hindi, Mandarin and Arabic.
• Good (10 to 20% WER): 18 languages, including Korean and Thai.
• Moderate (25 to 50% WER): 19 languages with noticeably higher error rates.

Beyond raw accuracy, Scribe's feature set is built for real business use, not just a basic transcript dump:

• Speaker diarization for up to 32 distinct speakers, useful for meeting recordings with several participants.
• Word-level timestamps on every transcription, which matters if you're generating captions or need to jump to a specific moment in a long recording. Making subtitles and SRT files puts them to use.
• Entity detection across 65 entity types, names, numbers, and in the medical-focused variant, clinical terms.
• Keyterm prompting, letting you supply up to 1,000 custom terms (brand names, jargon, acronyms specific to your business) to bias the transcription toward recognizing them correctly. Stopping transcripts from getting technical terms wrong has more tips.
• Multichannel support for up to 5 independent audio channels in a single file, and transcript editing through natural language instructions after the fact.

There's also a Scribe v2 Medical variant, which ElevenLabs states produces 35% fewer transcription errors on clinical audio specifically, aimed at healthcare settings where medical terminology trips up general-purpose transcription tools.

Price

ElevenLabs bills speech to text based on the duration of audio submitted, not word count or character count the way text to speech is billed. Its documentation states this directly, billing is "calculated per hour of audio, with rates varying by tier and model," but doesn't publish an exact per-hour credit figure on the page I was able to access. Confirm the current rate on ElevenLabs' pricing page before transcribing a large backlog of calls or meetings, since the per-hour cost at scale is the number that actually matters for budgeting, not the headline plan price.

Compared to Whisper and others

Whisper, OpenAI's open-source speech recognition model, is the most common free alternative people compare Scribe against. The practical difference is less about raw transcription accuracy on clear, single-speaker English audio, where both perform well, and more about everything around the transcription: Whisper is a model you self-host or run through a third-party API, with no built-in speaker diarization, keyterm prompting or entity detection out of the box the way Scribe ships them. If you're comfortable running your own infrastructure and your use case is simple single-speaker transcription, Whisper is genuinely free to run. If you need multi-speaker meeting transcripts, custom term recognition for your industry, or a managed service you don't have to host yourself, Scribe's built-in feature set does more of that work for you directly. Cost is the other real variable in this comparison: Whisper's compute cost depends entirely on where and how you run it, your own server, a cloud GPU rental, or a third-party API wrapping the same model, while Scribe's cost is a predictable, published per-hour rate on your ElevenLabs plan. For a business that wants one predictable bill rather than managing infrastructure, that predictability is often worth more than the headline "free" of a self-hosted model once engineering time is factored in.

A practical use case: sales calls

A common business use for Scribe is transcribing recorded sales calls, for coaching, for note-taking without a human on the line manually typing, or for feeding a CRM. Diarization matters a lot here, since a sales call transcript is far more useful when it clearly separates what the rep said from what the prospect said, rather than one undifferentiated block of text. Keyterm prompting is worth setting up specifically for your product names and any competitor names your reps discuss, since these are exactly the kind of terms a general-purpose transcription model is most likely to mishear or misspell on a first pass. Word-level timestamps then let you, or whoever reviews the call, jump directly to the specific moment a pricing objection or a competitor mention came up, instead of scrubbing through the full recording. Transcribing every call automatically covers the automatic setup.

What to check before relying on a transcript

• Spot-check against the actual audio, especially for anything moderate-accuracy-tier. If you're transcribing a language in Scribe's "moderate" accuracy tier (25 to 50% word error rate), treat the output as a rough draft that needs a human pass, not a finished transcript you can act on directly.
• Audio quality still matters, even with a strong model. A call recorded over a poor phone connection, with heavy background noise, or with two people talking over each other, will produce a noticeably worse transcript than clean, single-speaker audio, regardless of how accurate the underlying model is on paper. Cleaning up audio before transcribing it shows the quick fixes.
• Don't skip reviewing entity detection on anything with real numbers or names that matter. A misheard number in a transcribed contract discussion or a misspelled client name carried into a CRM record is a small transcription error with a real downstream cost.

A note on privacy for recorded calls

Transcribing calls or meetings at any real volume usually touches customer or employee conversations, which means privacy and consent considerations sit alongside the technical setup, not after it. Many places require some form of notice or consent before recording a call at all, separate from anything to do with the transcription tool itself. That's a legal question for your own jurisdiction and business, not something ElevenLabs' transcription accuracy has any bearing on, so confirm your recording and consent practices are sound before you scale up a call transcription workflow, regardless of which transcription tool sits downstream of that recording.

How I'd use this for a client

When a client needs call or meeting transcription built into a workflow, logging sales calls, generating meeting notes, feeding a support ticket system, I usually reach for a managed tool like Scribe specifically because the diarization and keyterm features handle real multi-speaker business audio out of the box, rather than building that layer myself on top of a bare model. Get in touch if you want transcription built properly into a larger workflow, not just a one-off upload.