What ElevenLabs released

On 10 November 2026, ElevenLabs announced two speech-to-text models. Scribe v3 (scribe_v3) transcribes recorded files through the batch endpoint. Scribe v3 Live (scribe_v3_live) transcribes streaming audio over a WebSocket. The pair replace Scribe v2 and Scribe v2 Realtime as the company's top transcription models. According to the models documentation, upgrading means changing the model_id value, and existing parameters keep working. The older models haven't been retired: the deprecation table lists only scribe_v1, so a team can run v2 and v3 side by side.

The price is the most useful part of the launch for builders. On the ElevenLabs API pricing page, Scribe v3 costs US$0.20 per hour of audio, the same as Scribe v2, and Scribe v3 Live costs US$0.22 per hour, the same as Scribe v2 Realtime. A team that already transcribes with v2 can test v3 on its own recordings without paying more per hour, and then decide based on its own results.

The benchmark claims and their caveats

Word error rate (WER) is the share of words a transcript gets wrong compared with a human reference transcript. ElevenLabs says Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages on an internal version of FLEURS, which is a benchmark of short sentences read aloud. The models documentation gives the absolute numbers: average WER drops from 16.94% to 12.38%. The launch post breaks this down by language, from a 23% reduction for English to 42% for German.

For Scribe v3 Live, the launch post reports a 43% reduction in semantic WER compared with Scribe v2 Realtime. Semantic WER counts only the errors that change a sentence's meaning. Plain WER falls by 31%. The documentation gives a slightly different semantic figure, 44% across 73 languages, so the exact number depends on which ElevenLabs page you read. The documentation also reports results on Earnings-22, a set of company earnings-call recordings, where Scribe v3 Live's WER falls from 5.99% to 3.91%.

These numbers need to be read with their limits in mind. FLEURS is clean read speech, which sounds very different from a noisy call centre or a meeting where people talk over each other. ElevenLabs scores each clip, averages the clips within each language, and weights every language equally. The documentation says this method can't be compared directly with the pooled, corpus-level WER that other vendors publish. Every accuracy figure comes from ElevenLabs' own tests, and AlphaSignal's coverage notes that none of the figures has been independently verified. ElevenLabs built the 16-language dataset behind its speaker-labelling claims itself. The company says it'll open-source the dataset "soon" but hasn't given a date.

Speaker labels, turn prediction and latency

Diarization means working out which speaker said each part of a recording. Scribe v3 labels up to 32 speakers in a batch file. The launch post reports 12% fewer diarization errors than Scribe v2, and 17% fewer on recordings with three or more speakers. Scribe v3 Live is the first Scribe model that labels speakers during a live stream, switched on with diarize=true. ElevenLabs reports a 3.1% diarization error rate for Scribe v3 Live on the English part of its dataset, against 5.9% to 7.3% for the other realtime models it tested. ElevenLabs scored the realtime results after aligning transcripts to the reference, so those numbers can't be compared with the batch numbers.

Turn prediction will matter most to people building voice agents. With commit_strategy=turn_prediction, Scribe v3 Live decides that a speaker has finished by looking at the conversation as well as the silence. A voice agent that waits only for silence feels slow, and one that guesses too early talks over the caller. The live model also adds multilang_mode=true for speakers who switch languages mid-sentence, and a keyterm library of up to 10,000 domain terms. The documentation warns that the official SDKs don't yet support turn prediction, multilingual mode or the keyterm library. Teams that want those features have to work directly with the WebSocket API for now.

ElevenLabs hasn't published a clear latency figure. The launch post gives no number. The realtime product page claims 150 ms in one section and under 100 ms in another, and says in neither place whether the figure is a median or a worst case. Teams should measure latency themselves, from their own hosting region.

Costs beyond the headline rate

The base hourly rate isn't the whole bill. The pricing page lists these add-ons, per hour of audio:

  • Keyterm prompting: US$0.05 on Scribe v3, US$0.08 on Scribe v3 Live
  • Keyterm library (Scribe v3 Live only): US$0.08
  • Transcript editing: US$0.07 on Scribe v3, US$0.12 on Scribe v3 Live

On those listed prices, a live voice agent that uses the keyterm library pays US$0.30 per hour of audio. The speech-to-text documentation says silence is billed too, because metering covers every second of audio sent. A streaming agent that keeps its connection open during long pauses pays for those pauses. Subscription plans include some transcription hours. The pricing page shows the US$22-a-month Creator plan with 110 hours of Scribe v3.

The same documentation sets a batch limit of 5 GB or 15 hours per file. Files longer than eight minutes are split into as many as four segments that are transcribed in parallel. Batch keyterm prompting accepts up to 1,000 terms of 50 characters each.

What Scribe v3 means for New Zealand builders

The speech-to-text documentation lists Māori among the supported languages, in the "Good" accuracy tier, which means a WER above 10% and up to 20%. That tier was measured on Scribe v2. ElevenLabs hasn't published a Scribe v3 figure for te reo Māori, and the launch post doesn't break out New Zealand English either. Builders working with te reo Māori, or with heavy code-switching between English and Māori, should run their own tests rather than take the 27% average at face value. Adding Māori place names and iwi names as keyterms is a cheap way to see whether keyterm prompting closes part of the gap.

Data location is the other question for New Zealand teams. The launch post says audio is processed in the United States by default. Enterprise customers can choose isolated environments in the EU, India or Singapore, and no Australian or New Zealand region is listed. Zero Retention Mode deletes the audio and transcript once a request finishes, and both models are available for private deployment. Teams that transcribe health consultations or customer calls should include US or Singapore processing in their privacy assessment before going live. ElevenLabs is also releasing a version of Scribe v3 Live tuned for medical conversations to healthcare teams in ElevenAgents.

A sensible first step is a small test set of your own recordings with checked reference transcripts. Include local accents, background noise and te reo Māori terms. Run Scribe v2 and Scribe v3 on the same files, compare WER and speaker labels, and add realtime latency measurements if you're building a voice agent. Members who build a repeatable evaluation harness for New Zealand speech can share it through ASAI projects, since vendor benchmarks don't measure local speech.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member