LATEST
AI Tools

ElevenLabs Scribe v3: 27% Fewer Errors, Same $0.20 an Hour

Updated November 15, 2026
13 min read
Editor's Note: This article was human-reviewed for accuracy
Silver dynamic microphone on a black stand against a dark background

Quick Answer & Key Takeaways

  • • ElevenLabs released Scribe v3 and Scribe v3 Live on November 10, 2026, claiming 27% fewer word errors on recorded audio and 43% fewer on live audio.
  • • API prices are unchanged at $0.20 an hour for Scribe v3 and $0.22 an hour for Scribe v3 Live, while a medical version tuned for clinical conversations arrived alongside the launch.
  • • Both models are HIPAA-eligible with a business associate agreement, and the accuracy figures come from ElevenLabs' own tests, not independent benchmarks.

ElevenLabs has a new answer to the oldest complaint about transcription: the transcript is almost right, except for the names, the numbers, and who said what. On November 10, 2026, the company released Scribe v3 for recorded audio and Scribe v3 Live for real-time audio, claiming 27% fewer word errors than Scribe v2 across 102 languages, and 43% fewer errors on the live model. The part that matters most for anyone paying the bill is what did not change: the API price stayed at $0.20 an hour for Scribe v3 and $0.22 an hour for Scribe v3 Live. This review checks what the two models do, what they cost through the API and the app, where the accuracy numbers come from, and the medical version that arrived with them.

What ElevenLabs launched on November 10

Scribe v3 and Scribe v3 Live are two models, not one, and they solve different problems. Scribe v3 transcribes recorded files after the fact: podcasts, videos, meetings, interviews. Scribe v3 Live transcribes audio as it arrives, for voice agents, live calls, and captions. Both are in the ElevenLabs API under the model IDs scribe_v3 and scribe_v3_live, and both are now the default transcription models in the ElevenLabs app. Switching from the previous generation is a one-line change of the model ID, and Scribe v2 and Scribe v2 Realtime remain available, according to ElevenLabs' documentation.

The live model carries the bigger changes. It labels speakers in real time, a first for the Scribe line, and it decides when a caller has finished speaking from the conversation itself rather than from silence alone, so a caller who pauses halfway through reading out a number is not cut off. It also handles speakers who switch languages mid-sentence, and it adds a separately priced keyterm library for terms that must be spelled a certain way. The recorded model handles up to 32 speakers in one file, and both models add controls for number formatting and removing filler words. One small date wrinkle: ElevenLabs' developer changelog lists the release under November 9, while the blog post is dated November 10. It is the same launch, recorded on two pages a day apart.

The launch sits inside a bigger week for the company: at its New York Summit on November 11, ElevenLabs announced 13 products and upgrades, including the Interaction Engine for routing voice-agent conversations and an expanded partnership with Nvidia. Scribe v3 is the piece of that week you can use today.

The accuracy claims, in ElevenLabs' own numbers

Every accuracy figure in this section comes from ElevenLabs testing its own models. That is not a dismissal; it is the fact that decides how much weight the numbers carry. The headline claims are measured on an internal version of FLEURS, a set of short, clean read-speech clips in 102 languages, and on a dataset of real meetings, agent calls, and voice notes in 16 languages that ElevenLabs built itself and has not published yet. The company says it will open-source that dataset soon. Until it does, no outside lab can repeat the speaker-label scores.

ClaimWhat ElevenLabs reportsMeasured on
Scribe v3 word errors vs Scribe v227% fewer; word error rate fell from 16.94% to 12.38%Internal FLEURS, 102 languages
Scribe v3 Live semantic errors vs Scribe v2 Realtime43% fewer; plain word error rate down 31%Announcement; docs say 44% across 73 languages
Scribe v3 speaker-label errors vs Scribe v212% fewer overall; 17% fewer on recordings with 3+ speakersElevenLabs' own 16-language dataset
Scribe v3 Live speaker-label errors39% fewer than the next-best realtime model tested; 3.1% diarization error rate on EnglishSame dataset; rival models scored 5.9% to 7.3%
By language (recorded)German 42%, Spanish 35%, English 23%, Swahili 34%, Tamil 29%, Thai 28%Internal FLEURS
By language (live)Greek 77%, Danish 54%, Croatian 45%, Hindi 36%ElevenLabs' own tests

Two numbers deserve a second look. First, the live model's 43% figure counts semantic word errors, which only count an error when the meaning changes; on the plain word error rate the improvement is 31%. Second, the announcement says 43% across 102 languages while the documentation says 44% lower semantic word error rate across 73 languages. The pages do not explain the difference, so treat the headline as roughly a 43 to 44% improvement on the vendor's own measure, not a precise figure.

There is one outside test so far. Voice Arena, an account on X that publishes speech-to-text test results, ran Scribe v3 Live on more than 70 hours of real, unscripted phone calls and posted the results on November 11. It ranked the model first among streaming models in Romanian and Indian English, and third in US English, where AssemblyAI's Universal-3.6 Pro Realtime leads. On the same calls, the error score fell 66% in Romanian and 18% in US English against Scribe v2 Realtime. Voice Arena's operator and scoring method are not public, so read it as a directional check, not a verdict. On the Artificial Analysis board for recorded audio, Scribe v3 sat second of 64 models at a 1.9% word error rate when read on November 12, behind Alibaba Cloud's Fun-Realtime-ASR-preview at 1.7%, with Scribe v2 fifth at 2.2%.

One customer result comes from the launch itself. Dutch energy supplier Eneco, which serves more than 2 million homes and businesses, tested Scribe v3 Live on its voice agent before launch. ElevenLabs says the agent recognised postcodes correctly on the first try 40% more often, and house numbers 15% more often. That is a company-reported result from a single deployment, and it is the kind of test that matters more than a benchmark: real callers, real addresses, no clean read speech.

What it actually costs

The API price is the simple part. Scribe v3 costs $0.20 per hour of audio and Scribe v3 Live costs $0.22, the same rates as the models they replace. At November 15, 2026 market rates of 1 USD to 0.7566 GBP and 1.4271 CAD, that is about £0.15 and C$0.29 an hour for the recorded model, and about £0.17 and C$0.31 for the live model. Treat the conversions as approximations; ElevenLabs bills in US dollars.

API chargeScribe v3Scribe v3 Live
Base rate, per hour of audio$0.20$0.22
Keyterm prompting, per hour+$0.05+$0.08
Experimental transcript editing (recorded files)+$0.07Not offered
100 hours of audio, base rate$20.00$22.00
100 hours with keyterms on$25.00$30.00

The app is where the same transcription costs far more. In the ElevenLabs app, speech to text uses about 330 credits a minute, drawn from the same monthly credit pool as text to speech, music, and dubbing. A weekly one-hour podcast, about 4.3 hours of audio a month, costs about $0.86 through the API, or about 85,000 credits in the app. That fits inside the $22 Creator plan's 121,000 credits but uses about 70% of them, and it does not fit in the $6 Starter plan's 30,000 credits at all. The free plan's 10,000 credits cover roughly 30 minutes of transcription a month.

Plan (monthly billing)PriceCreditsApp transcription if you spent every creditAPI hours listed
Free$010,000About 30 minutesAbout 4.5 hours
Starter$6 ($1 first month)30,000About 90 minutes30 hours
Creator$22 ($11 first month)121,000About 6 hours110 hours
Pro$99600,000About 30 hours495 hours

Three pieces of fine print change the math. First, the first-month prices renew at the full rate, so Starter is $6 and Creator is $22 from month two. Second, unused paid credits roll over for up to two months while you stay subscribed. Third, ElevenLabs bills every second you send, silence included, so long dead air at the start and end of a file is billed like speech. Trim it before you upload. Annual billing lowers the monthly equivalent to $5 for Starter and $18.33 for Creator, with Pro at $82.50, Scale at $249.17, and Business at $825.

Scribe v3 Medical: the healthcare version

Announced on November 13, 2026, Scribe v3 Medical is a version of Scribe v3 Live tuned for clinical conversations, built to recognise thousands of medical terms that a general model mangles: drug names, procedure names, dosages. It runs inside ElevenAgents, ElevenLabs' voice-agent product, rather than as a standalone API model, and a financial-services version is in development. The November 10 launch post had already described a medical-tuned version of the live model, so the November 13 announcement is the formal product, not a surprise addition.

The healthcare pitch rests on two compliance facts. Both Scribe v3 models are HIPAA-eligible with a business associate agreement, and ElevenLabs says they are covered by its SOC 2, ISO 27001, and PCI DSS Level 1 programs. Data runs in the US by default, with isolated environments in the EU, India, and Singapore for enterprise customers. Note the limit: the isolated environments and data-residency options are enterprise features, not something the $6 plan buys.

The clearest real deployment is in Poland, where ElevenLabs and Nvidia power a voice agent for the country's central medical appointment system, calling patients who only have landlines in a system handling about 40 million appointments a year with no-show rates of 10 to 20 percent. The two companies also announced plans to add 15 megawatts of Nvidia Vera Rubin computing capacity. None of that changes what Scribe v3 costs, but it shows where the medical model is aimed: national-scale appointment systems, not individual clinics.

Scribe v3 vs the alternatives

Speech to text is crowded right now, and Scribe v3's price is not the lowest on the market. The comparison that matters is accuracy per dollar, and on that measure the honest answer is that the independent evidence is still thin. What can be compared today is price, availability, and the one outside test that exists.

ModelBest forList priceLive speaker labels
ElevenLabs Scribe v3Recorded audio, interviews, podcasts$0.20 per hourNo (recorded model)
ElevenLabs Scribe v3 LiveVoice agents, live calls$0.22 per hourYes
Soniox STT v6Voice agents on a budget$0.12 per hourYes
Mistral Voxtral 3Phone agents, telephony$0.04 per minute ($2.40 per hour)Yes
NVIDIA Nemotron 3.5 TranscribeFree transcription with speaker labelsFreeYes
OpenAI gpt-4o-transcribe-miniDevelopers already on OpenAIAbout $0.18 per hour (OpenAI's estimate)No

Soniox STT v6, reviewed here a day before this article, is the sharpest price comparison: $0.12 an hour for real-time use against Scribe v3 Live at $0.22, nearly half the price. Voice Arena's phone-call test ranked Scribe v3 Live first in Romanian and Indian English, but that test did not run Soniox, so the accuracy gap between the two is unmeasured. If your calls are mostly in English and your budget is tight, Soniox is the cheaper call; if your audio is multilingual or messy, Scribe v3's 102-language coverage and diarization scores are the argument for paying more. Mistral's Voxtral 3 is far more expensive per hour at $2.40, but it is a full voice-agent model with telephony built in, not a transcription-only service. Nemotron 3.5 Transcribe is free, which makes it the right first test for anyone who has not settled on a vendor. And OpenAI's smallest transcription model lists lower than Scribe v3 at about $0.18 an hour, though OpenAI labels its per-minute figures as estimates of token billing rather than a flat rate.

There is also a speed trade-off that the price table hides. Independent benchmark coverage of the launch notes that Scribe v3 runs slower and costs roughly twice as much as comparable models from Microsoft and xAI, which process audio many times faster. ElevenLabs is selling accuracy, not speed, and for a recorded podcast that trade is invisible. For a voice agent that must answer in under a second, it is the first thing to test.

How to try it

If you already use ElevenLabs, Scribe v3 is already in your account as the default transcriber, so the fastest test costs nothing extra. Upload one short recording in the app, check the credit hit against the 330-credits-a-minute figure, and read the transcript for the names and numbers in your own audio before you trust it with a full season.

For the API, the switch from Scribe v2 is a change of the model ID to scribe_v3 for recorded audio or scribe_v3_live for streaming, with existing request parameters working without changes. Three features need a direct WebSocket connection because the official SDKs do not support them yet: turn prediction, multilingual mode, and the keyterm library. Files up to 5 GB and 15 hours go through the API. If you transcribe a lot, use the API rather than the app: the same hour costs $0.20 through the API and a large share of a monthly credit pool in the app.

What to watch before you switch

The most common mistake is comparing the app price with the API price as if they were the same product. They are not. The app charges credits from a shared pool, so a heavy transcription month leaves nothing for voice work on the same plan. Run one short test file first, then check your usage page against the math in the pricing section before uploading a whole archive.

The second mistake is trusting the benchmark numbers on your own audio. ElevenLabs' figures come from short, clean read speech and its own curated dataset. A podcast recorded in a kitchen, a call with crosstalk, or a speaker with a heavy accent will score differently, and the one outside test found the gains much smaller in US English, 18%, than in Romanian, 66%. Test on your worst audio, not your best.

The third is billing. When ElevenLabs launched the previous realtime model, a developer reported on the company's subreddit that the API cost looked ten times higher than expected, as if billed at the app rate instead of the API rate. Separately, ElevenLabs' status page logged a short incident in January 2026 in which some speech-to-text requests were overcharged and later credited back. Neither incident proves a current problem, but both point to the same habit: check the first bill line by line.

FAQ

How much does ElevenLabs Scribe v3 cost?

Scribe v3 costs $0.20 per hour of audio through the API, and Scribe v3 Live costs $0.22 per hour, the same rates as the models they replace. Keyterm prompting adds $0.05 an hour on the recorded model and $0.08 on the live model. In the ElevenLabs app, transcription uses about 330 credits a minute from your plan's shared credit pool.

What is the difference between Scribe v3 and Scribe v3 Live?

Scribe v3 transcribes recorded files after the fact, while Scribe v3 Live transcribes audio as it arrives, for voice agents and live calls. The live model adds real-time speaker labels, conversation-aware turn detection, and a multilingual mode for speakers who switch languages mid-sentence.

Is Scribe v3 more accurate than Scribe v2?

ElevenLabs says it makes 27% fewer word errors across 102 languages, measured on its own internal version of the FLEURS benchmark. The one outside test so far, Voice Arena's phone-call test, found the live model first in Romanian and Indian English and third in US English. Test it on your own audio before trusting the number.

What is Scribe v3 Medical?

Scribe v3 Medical is a version of Scribe v3 Live tuned for clinical conversations, announced on November 13, 2026. It recognises medical terminology and runs inside ElevenAgents for healthcare teams. Both Scribe v3 models are HIPAA-eligible with a business associate agreement.

Is Scribe v3 free to try?

Yes, in a limited way. Speech to text is on the free plan, which includes 10,000 credits a month, enough for roughly 30 minutes of transcription in the app. Through the API, the free plan covers about 4.5 hours of audio, according to ElevenLabs' API page.

Should I use the API or the app?

Use the API if you transcribe more than a few hours a month: the same hour costs $0.20 through the API and a large share of a monthly credit pool in the app. Use the app if you already pay for ElevenLabs voice and only transcribe occasionally, and watch the credit balance.

Scribe v3 is a real improvement at the same price, and that is the whole story in one sentence. The accuracy gains are vendor-measured, the independent evidence is one phone-call test and a leaderboard, and the app pricing punishes anyone who transcribes in bulk. If you already use ElevenLabs, switch the model ID and test it on your hardest audio. If you are choosing a transcription vendor from scratch, run Scribe v3 against Soniox STT v6 and the free Nemotron 3.5 Transcribe on the same recordings before you commit. Read more of our speech-to-text coverage in the Soniox STT v6 review and the ElevenAI explainer, which uses Scribe v3 for meeting notes.

Sources

13 min read
22 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more