Scribe v3 by ElevenLabs: Speech to text with 27% fewer errors in 90+ languages
ElevenLabs released two speech-to-text models on November 10, 2026. Scribe v3 is aimed at recorded audio, and Scribe v3 Live is aimed at real-time use such as voice agents and live captions. ElevenLabs reports that Scribe v3 cuts word errors by 27% against Scribe v2 on the FLEURS benchmark, where its average word error rate fell from 16.94% to 12.38% across 102 languages. Product Hunt's listing frames the release around 90-plus languages, the language coverage ElevenLabs documents.
Scribe v3 labels up to 32 speakers and can be steered toward custom vocabulary. The API accepts keyterm prompts of up to 1,000 terms, while a separate key term library holds up to 10,000 terms. Other additions include digit or word number formatting, secondary language codes and edits written as plain English rules. On the live side, Scribe v3 Live adds realtime speaker labels, turn detection that reads conversation context instead of silence alone, and a median latency of 150 ms. ElevenLabs puts Live's semantic word error rate 44% below Scribe v2 Realtime on FLEURS.
Prices held at $0.20 per hour for batch audio and $0.22 per hour for live audio, the same as the models they replace. Scribe v3 runs in the US by default, with isolated EU, India and Singapore environments for enterprises. Scribe v3 Live is available in ElevenAgents, and both models are in the ElevenAPI.
Why it matters
For teams building voice products, the main change is accuracy without a price increase. Speaker labels, keyterm biasing and live diarization are features that could otherwise need custom engineering, so they may shorten work for call analytics, meeting tools and legal or medical transcription. Founders switching from a rival API should check whether the cost per hour and the 150 ms latency hold up in their own stack.
For investors and sellers, the signal is that a large model vendor is folding diarization, live transcription and regional hosting into one API. That narrows room for single-feature transcription startups to compete on accuracy alone. Buyers will still want independent tests, since the headline gains come from the vendor's own benchmarks.
Key points
- 1Scribe v3 makes 27% fewer word errors than Scribe v2 on the FLEURS benchmark across 102 languages.
- 2Scribe v3 labels up to 32 speakers and is the default model in the ElevenLabs app.
- 3Scribe v3 Live has a median latency of 150 ms and is available in ElevenAgents and the ElevenAPI.
- 4Batch pricing is $0.20 per hour and live pricing is $0.22 per hour, unchanged from the predecessors.
- 5Keyterm prompting accepts up to 1,000 terms per request, and a separate library stores up to 10,000.
