ElevenLabs released Scribe v3 and Scribe v3 Live on November 10, 2026. The most useful change for application developers is the separation between a better recorded-audio workflow and new controls for live conversations. A model upgrade alone will not enable every streaming feature.
Two models, two integration paths
The recorded-audio model uses scribe_v3. It adds options for formatting numbers and supplying secondary-language hints. The streaming model, scribe_v3_live, adds speaker labeling, multilingual mode and context-based turn prediction.
Speaker labeling answers “who said this?” Turn prediction addresses a different problem: when an application should treat the speaker’s thought as complete. In a voice interface, a premature handoff can cut off a question; an overly patient one makes the conversation feel slow. Those are separate acceptance tests, even when the transcript looks correct.
Check the streaming controls before switching
The current model documentation says Live turn prediction, multilingual mode and the keyterm library require a direct WebSocket integration because official SDK support for those controls is not yet available. It also says the default commit strategy remains manual. Developers seeking context-based turn prediction therefore need an explicit configuration change, not just a different model name.
For a staged evaluation, keep the old model as a comparison and use recordings you are authorized to process. Include overlapping speakers, pauses, names, numbers and code-switching. Measure speaker attribution and end-of-turn behavior separately from transcription errors. Do not send sensitive recordings merely to try the feature.
English and Hebrew: support is not a local benchmark
ElevenLabs lists both English and Hebrew among its supported languages. Mixed-language transcription preserves the language being spoken rather than translating the conversation. That is relevant to teams that alternate between Hebrew discussion and English product names or technical terms.
We did not verify a Hebrew-specific Scribe v3 accuracy result. The language-quality tiers currently shown in the documentation describe v2, so they should not be presented as a v3 scorecard.
Read the accuracy claims carefully
The launch article and current model documentation give different figures and language counts for the Live accuracy improvement. SysSeder is therefore not treating either percentage as a settled comparison. These are also vendor-reported evaluations, not our own testing.
Bottom line: Scribe v3 is worth evaluating for richer transcription workflows. For live applications, first verify which controls your integration actually exposes, then test conversational behavior on representative audio. Vocabulary-library use can add charges; this report did not run paid tests.
Sources
- ElevenLabs launch announcement, November 10
- Model documentation and migration details
- Speech-to-text language and feature documentation
Source check: November 12, 2026. Feature availability and documentation can change.