ElevenLabs Scribe v3 cuts transcription errors 27% at $0.20 an hour
ElevenLabs released Scribe v3 and Scribe v3 Live on November 10, 2026, two new speech-to-text models that replace Scribe v2 and Scribe v2 Realtime as the company's most accurate transcription engines. According to ElevenLabs, Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, while Scribe v3 Live cuts semantic word error rate by 43% compared with Scribe v2 Realtime. Both models are available in ElevenAPI under the IDs `scribe_v3` and `scribe_v3_live`, and Scribe v3 is now the default transcription model inside the ElevenLabs app.
The price did not move. The official API pricing page lists Scribe v3 at $0.20 per hour of audio, the same rate as Scribe v2 and Scribe v2 Medical, and Scribe v3 Live at $0.22 per hour, the same rate as Scribe v2 Realtime. Developers who already run Scribe v2 in production therefore get the accuracy gains without a change in base cost, and ElevenLabs says the switch requires only a new model ID. Add-ons such as keyterm prompting and natural-language transcript editing are billed on top of the base rate, and silence counts toward billed duration.
Scribe v3 Live also brings features that were missing from the previous realtime model: speaker diarization in realtime, a multilingual mode that follows language switches in the middle of a sentence, and turn detection that uses conversation context instead of silence alone. ElevenLabs says this is the same turn detection that already runs inside its ElevenAgents voice-agent platform, now exposed directly in the API. Neither Scribe v2 nor Scribe v2 Realtime has been deprecated, and the model documentation lists both as superseded but still available.
··········
ELEVENLABS SCRIBE V3 AND SCRIBE V3 LIVE: WHAT WAS RELEASED.
The release consists of two general models and one specialized variant. Scribe v3 is the batch model for recorded audio, called through the standard `POST /v1/speech-to-text` endpoint. Scribe v3 Live is the streaming model for live audio, reached over the WebSocket endpoint `wss://api.elevenlabs.io/v1/speech-to-text/realtime` with `scribe_v3_live` passed as the model ID. ElevenLabs also announced a version of Scribe v3 Live tuned for medical conversations, offered to healthcare teams inside ElevenAgents, without publishing a separate name or API identifier for it.
Both general models support more than 90 languages according to the developer documentation, and the accuracy evaluation covers 102 languages. For recorded files, the documentation sets a maximum upload of 5 GB and a maximum duration of 15 hours, with multichannel mode accepting up to five channels whose combined length must stay under 15 hours. Batch jobs longer than eight minutes are split into up to four segments that are transcribed in parallel, so a 15-minute file runs with a concurrency of two and a two-hour recording with a concurrency of four.
The feature list for Scribe v3 includes speaker diarization for up to 32 speakers, keyterm prompting for up to 1,000 terms of up to 50 characters each, a numbers format option that writes figures as digits or as words, secondary language codes for mixed-language recordings, word-level or character-level timestamps, audio event tagging for sounds such as laughter or applause, entity detection, and transcript editing. Transcript editing accepts formatting rules written in plain English; the example ElevenLabs gives is an instruction to write every date as YYYY-MM-DD. Output can be exported as SRT subtitles, DOCX or PDF, and asynchronous results can be delivered by webhook.
··········
SCRIBE V3 ACCURACY: WER, DIARIZATION AND PER-LANGUAGE GAINS.
ElevenLabs measured multilingual accuracy on an internal version of the FLEURS benchmark, a set of short, clean read-speech clips in 102 languages. Each language is scored separately and then averaged, so a low-resource language weighs as much as English in the headline figure. On that basis the documentation reports an average word error rate of 16.94% for Scribe v2 and 12.38% for Scribe v3, the reduction behind the 27% claim. ElevenLabs warns that it computes WER per clip before averaging, so these numbers cannot be compared directly with corpus-level WER figures published by other vendors.
For Scribe v3 Live, the announcement reports a 43% reduction in semantic WER against Scribe v2 Realtime and a 31% reduction in standard WER. Semantic WER counts only errors that change the meaning of the transcript, which makes it a more lenient metric than standard WER. The model documentation gives a slightly different figure, 44% lower semantic WER on FLEURS across 73 languages, together with results on two public English datasets: on Earnings-22, a set of earnings-call recordings, WER falls from 5.99% to 3.91%, and on VoxPopuli, recorded parliamentary speech, from 2.13% to 1.71%.
........
Benchmark (ElevenLabs figures) | Previous model | Scribe v3 / v3 Live | Change |
FLEURS, 102 languages, WER (batch) | 16.94% (Scribe v2) | 12.38% (Scribe v3) | 27% fewer errors |
FLEURS, semantic WER (live) | Scribe v2 Realtime | Scribe v3 Live | 43% lower (44% in docs) |
FLEURS, standard WER (live) | Scribe v2 Realtime | Scribe v3 Live | 31% lower |
Earnings-22, WER (live) | 5.99% | 3.91% | 35% fewer errors |
VoxPopuli, WER (live) | 2.13% | 1.71% | 20% fewer errors |
Diarization errors (batch) | Scribe v2 | Scribe v3 | 12% fewer, 17% with 3+ speakers |
Realtime diarization, English DER | 5.9% to 7.3% (other realtime models) | 3.1% | 39% fewer than next best |
........
The per-language results show where the batch model improved most. Against Scribe v2, Scribe v3 lowers WER by 42% in German, 35% in Spanish and 23% in English, and by 34% in Swahili, 29% in Tamil and 28% in Thai. On the live side, the largest gains over Scribe v2 Realtime are 77% in Greek, 54% in Danish, 45% in Croatian and 36% in Hindi. ElevenLabs states that every language it tested improved for both models.
Diarization, the task of labeling who is speaking, was evaluated on a dataset ElevenLabs built from real meetings, agent calls and voice notes in 16 languages, which the company says it plans to open-source. Scribe v3 produces 12% fewer diarization errors than Scribe v2 overall and 17% fewer on recordings with three or more speakers. For the live model, ElevenLabs reports a diarization error rate of 3.1% on the English portion of the dataset against 5.9% to 7.3% for other realtime models, scored after aligning each transcript to the reference, a method the company says makes those results non-comparable with batch DER.
The comparisons with other vendors are less specific. ElevenLabs says Scribe v3 has fewer word errors than every other batch model it tested on the languages both models support, and that Scribe v3 Live beats every other realtime model it tested, but it does not name the competing systems or publish their scores. Until the diarization dataset is released and third parties rerun the tests, these competitive claims remain vendor results. The only customer figure in the announcement comes from Eneco, a Dutch energy provider, which reported that Dutch postcodes were recognized correctly on the first attempt 40% more often and house numbers 15% more often.
··········
SCRIBE V3 LIVE: TURN PREDICTION, REALTIME DIARIZATION AND MULTILINGUAL MODE.
The realtime model changes more than accuracy. Turn prediction, enabled with `commit_strategy=turn_prediction`, decides when a speaker has finished by looking at conversation context and audio together, which addresses one of the most common complaints about voice agents: replies that start while the caller is pausing mid-thought, or long gaps after the caller has finished. The default commit strategy remains manual, so existing integrations keep their current behavior until developers opt in.
Realtime diarization is new for the streaming model and is switched on with `diarize=true`, so live transcripts can label speakers as a meeting or call unfolds. Multilingual mode, enabled with `multilang_mode=true`, follows speakers who change language in the middle of a sentence, a frequent pattern in customer service across India, Southeast Asia and parts of Europe. The live model also supports entity detection across 65 entity types with optional redaction, background audio filtering, word-level timestamps, PCM audio from 8 kHz to 48 kHz and μ-law encoding for telephony.
Vocabulary control works at two levels. A standard realtime request accepts up to 50 keyterms of up to 20 characters each, and Scribe v3 Live adds a keyterm library meant for large product catalogs, drug names or account terminology. The documentation states that a library can contain up to 10,000 terms, although another line on the same page mentions 100,000, so developers planning very large libraries should confirm the limit with ElevenLabs. The documentation also notes that the official SDKs do not yet support turn prediction, multilingual mode or the keyterm library, which means teams that want those features at launch need to call the WebSocket API directly.
··········
SCRIBE V3 PRICING, PLAN ALLOWANCES AND CONCURRENCY LIMITS.
Prices below are as listed on the official ElevenLabs API pricing page on November 11, 2026. The base rate is identical on every plan, and the add-ons are charged per hour of audio on top of it.
........
Model | Base price | Keyterm prompting | Keyterm library | Transcript editing |
Scribe v3 (batch) | $0.20 per hour | +$0.05 per hour | n/a | +$0.07 per hour |
Scribe v2 (batch) | $0.20 per hour | +$0.05 per hour | n/a | +$0.07 per hour |
Scribe v3 Live | $0.22 per hour | +$0.08 per hour | +$0.08 per hour | +$0.12 per hour |
Scribe v2 Realtime | $0.22 per hour | +$0.08 per hour | n/a | +$0.12 per hour |
........
The plans differ in monthly price, in the audio volume their allowance covers and in how many transcription requests can run at once. The included hours on the pricing page correspond to each plan's monthly price divided by the hourly rate, which means the allowance is a monetary budget expressed in hours. Starter costs $6 a month and Creator $22 a month, with discounted first months of $1 and $11.
........
Plan | Monthly price | Scribe v3 hours | Scribe v3 Live hours | Batch concurrency | Realtime concurrency |
Free | $0 | 4.5 | 2.5 | 8 | 6 |
Starter | $6 | 30 | 27 | 12 | 9 |
Creator | $22 | 110 | 100 | 20 | 15 |
Pro | $99 | 495 | 450 | 40 | 30 |
Scale | $299 | 1,495 | 1,359 | 60 | 45 |
Business | $990 | 4,950 | 4,500 | 60 | 45 |
........
Enterprise contracts carry custom pricing and what ElevenLabs calls elevated concurrency. Diarization and entity detection have no separate price on the pricing page, and ElevenLabs has not published latency figures for either new model. The only latency number in the documentation belongs to Scribe v2 Realtime, with a median of about 150 milliseconds to a committed transcript, excluding application and network latency.
··········
[[ADPLUS_300x250_1]]
WHAT SCRIBE V3 COSTS FOR A CONTACT CENTER AND A VOICE AGENT.
Two workloads show how the add-ons shape the final bill. The first is a contact center that transcribes 2,000 hours of recorded calls a month with Scribe v3 for quality review. The second is a voice agent that streams 600 hours of live calls a month through Scribe v3 Live. In both cases the base rate is unchanged from the Scribe v2 generation, so the before and after cost is the same for an identical configuration, and the difference lies in error counts.
........
Workload | Configuration | Monthly cost |
2,000 h recorded calls, Scribe v3 | Base rate only | $400 |
2,000 h recorded calls, Scribe v3 | Base + keyterm prompting | $500 |
2,000 h recorded calls, Scribe v3 | Base + keyterms + transcript editing | $640 |
600 h live calls, Scribe v3 Live | Base rate only | $132 |
600 h live calls, Scribe v3 Live | Base + keyterm library | $180 |
600 h live calls, Scribe v3 Live | Base + keyterm library + transcript editing | $252 |
........
The add-ons can raise the batch bill by 60% and the live bill by about 91%, so teams should enable them only where custom vocabulary or automated formatting removes manual work. On the accuracy side, a rough estimate helps size the gain. At an assumed 150 words per minute, 2,000 hours of calls contain about 18 million words. Applying the Earnings-22 error rates that ElevenLabs reports for its live models, 5.99% and 3.91%, which are the only public-dataset WER values it gives, the number of word errors would fall from roughly 1.08 million to roughly 0.70 million a month. Real call audio is noisier than earnings calls, so actual error rates will be higher, but the relative reduction indicates how much less correction work a review team could face.
The price parity favors high-volume batch users most, because they gain accuracy across thousands of hours at no extra base cost and need no code changes beyond the model ID. Voice-agent builders gain the new turn prediction and realtime diarization, but they pay for the keyterm library separately and must bypass the official SDKs to use the new realtime options for now. Readers comparing speech engines can also look at our coverage of Microsoft's MAI-Transcribe-2-Streaming model, the most recent realtime transcription launch from a large cloud provider.
··········
MIGRATING FROM SCRIBE V2 TO SCRIBE V3 IN THE ELEVENLABS API.
For batch transcription, the upgrade consists of changing the `model_id` field in the request to the speech-to-text endpoint. The illustrative request below uses the endpoint and model ID from the ElevenLabs documentation; parameter names for optional features should be checked against the API reference before use.
# Illustrative: batch transcription with Scribe v3
curl -X POST "https://api.elevenlabs.io/v1/speech-to-text" \
-H "xi-api-key: $ELEVENLABS_API_KEY" \
-F "model_id=scribe_v3" \
-F "file=@call-recording.mp3"For streaming, the model ID goes in the query string of the WebSocket connection as `model_id=scribe_v3_live`. Since the default commit strategy is still manual, a Scribe v2 Realtime integration that swaps only the model ID keeps its current turn handling; turn prediction, realtime diarization and multilingual mode each need to be requested explicitly. Teams that rely on the ElevenLabs SDKs can migrate the model now and add those options later, once the SDKs support them.
In the ElevenLabs app, Scribe v3 is already the default, so creators who transcribe interviews or podcasts there receive the new model without any action. Inside ElevenAgents, Scribe v3 Live is available for voice agents, along with the medical variant for healthcare teams. ElevenLabs continues to expand its model range beyond speech recognition, as with the recent SFX v3 sound effects model in ElevenCreative and ElevenAPI.
··········
DATA RESIDENCY, ZERO RETENTION AND COMPLIANCE FOR SCRIBE V3.
Both models process audio in the United States by default, with isolated environments in the EU, India and Singapore available for customers that need data residency. A Zero Retention Mode deletes audio and text once a request completes, an option aimed at contact centers and healthcare providers that are not allowed to keep recordings on third-party systems. ElevenLabs lists SOC 2, ISO 27001 and PCI DSS Level 1 coverage, and says the service is HIPAA-eligible when the customer signs a Business Associate Agreement. Private deployment is offered for organizations that cannot use the shared cloud.
These controls are the same ones that apply across ElevenLabs' platform, and the announcement does not describe any compliance change specific to Scribe v3. For regulated buyers, the medical variant of Scribe v3 Live is the most relevant addition, although ElevenLabs has published no accuracy figures, pricing or separate model ID for it.
··········
ELEVENLABS SCRIBE V3: CONFIRMED FACTS AND OPEN QUESTIONS.
The verified part of the launch is clear. Scribe v3 and Scribe v3 Live are available now in ElevenAPI, Scribe v3 is the default in the ElevenLabs app, Scribe v3 Live runs in ElevenAgents, and both keep the hourly prices of the models they replace, $0.20 for batch and $0.22 for live audio. Migration is a model ID change, and Scribe v2 and Scribe v2 Realtime remain available with no announced retirement date.
The accuracy gains are ElevenLabs' own measurements: a 27% lower WER on its FLEURS evaluation, 43% or 44% lower semantic WER for the live model depending on the source page, and lower diarization error on a dataset that has not yet been released. The unnamed competitors, the missing latency figures for both new models, the unpriced medical variant and the inconsistent keyterm library limit are the points that remain open. Developers who need independent confirmation can run their own audio through both generations at the same price and compare error counts directly before switching production traffic.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
