ElevenLabs Launches Scribe v3 and v3 Live Transcription. How Good Is Its Arabic?

ElevenLabs has released two new speech-to-text models, Scribe v3 and Scribe v3 Live, in an announcement on 10 November 2026. While the release introduces real-time diarization and context-aware turn detection, Arabic users should test before switching: the vendor itself rates Arabic in its third accuracy tier, dialects get no mention, and there is no Middle East data region.

Conversational Intelligence and What Changed

The launch introduces two models: Scribe v3 for recorded audio and Scribe v3 Live for low-latency streaming. According to vendor claims, Scribe v3 achieves 27% fewer word errors than Scribe v2 across 102 languages on an internal FLEURS benchmark of clean read speech. ElevenLabs highlights error drops in German of 42%, Spanish of 35%, English of 23%, Swahili of 34%, Tamil of 29%, and Thai of 28%, while leaving Arabic unmentioned. Diarization errors on recordings with 3 or more speakers dropped 17% according to the official release.

For conversational voice bots, Scribe v3 Live tackles callers being cut off mid-sentence. ElevenLabs claims a 43% reduction in semantic word error rate over Scribe v2 Realtime, introducing real-time diarization and context-based turn detection that avoids interrupting callers who pause or speak digits slowly (model announcement). Controls include digits or words formatting, plain-English transcript editing, and keyterm prompting up to 1,000 terms in batch or 10,000 terms in real-time per the technical documentation. Both are now the default in the ElevenLabs app and available in the API as scribe_v3 and scribe_v3_live, with a medical variant in ElevenAgents.

Pricing and the Latency Trade-Off

According to the published API pricing, Scribe v3 costs $0.20 per audio hour, Scribe v3 Live costs $0.22 per hour, and keyterm prompting adds an extra $0.05 per hour, with the pricing page claiming support for 90+ languages.

Independent evaluation suggests precision carries performance trade-offs. As reported by The Decoder on 12 Nov 2026, benchmark testing by Artificial Analysis rates Scribe v3 among the most accurate models on recorded audio—surpassing models from Google, Mistral, OpenAI, Microsoft, and xAI, trailing only an Alibaba preview model. However, Scribe v3 lands mid-pack on speed and cost: Microsoft and xAI models process audio many times faster at slightly higher error rates and roughly half the price.

The Arabic Reality: Third-Tier Accuracy and Dialect Gaps

For Arabic speakers the picture is less clear. Per the vendor documentation, Arabic sits in the third tier labeled “Good” alongside Hebrew and Korean, behind English, German, and French (“Excellent”) and Hindi and Persian (“High Accuracy”). Regional dialects—such as Gulf, Egyptian, Levantine, or Maghrebi—are not mentioned anywhere in the documentation, so dialect accuracy is unknown. Note that this table was written for Scribe v2. It is worth comparing against Arabic-first tools such as Munsit speech recognition, as discussed in our Arabic transcription guide.

On the plus side, Scribe v3 Live adds a multilingual mode handling mid-sentence switching, useful for the Gulf Arabic-English code-switching. Data location is another question: processing runs in the United States by default, with isolated enterprise environments limited to the European Union, India, and Singapore, and no Middle East data region available (infrastructure details), although Zero Retention Mode deletes audio and text upon completion.

What to Do Now: Run a Controlled Test

Rather than adopting Scribe v3 outright for automated workflows or Arabic video subtitles, run a targeted evaluation on your own audio:

• Record 2-3 minutes of your local dialect mixed with common English business or technical terminology.
• Transcribe the file using Scribe v3 via the web application or API under default settings.
• Review the output, explicitly counting transcription errors across dialectal phrases, names, and spoken numbers.
• Repeat the test with keyterm prompting enabled to verify whether the extra $0.05 per hour resolves naming errors.