Automatic speech recognition now bundles much more than a text string. Scribe v2 offers word timing and speaker labels, but none of these replaces editorial verification of an interview.
Capabilities with distinct meanings
ElevenLabs documents Scribe v2 for more than 90 languages, word-level timestamps, diarization for up to 32 speakers and audio-event tagging. It also describes keyterm prompting and entity detection. Language count refers to recognition support, not a guarantee of equal word error rates; diarization labels different speakers but does not identify a person’s real-world name.
Choose batch or real time deliberately
The vendor documents Scribe v2 Realtime as a separate model for streaming use. Its published approximate latency excludes application and network delay. A journalist uploading an hour-long recording should favor a batch workflow with correction tools, while a contact-center agent needs provisional and final transcript handling. A partial live hypothesis can change later, so do not treat interim words as a legal record.
Handling sensitive entities
Entity detection is useful for finding names, payment card references and other categories, yet it may introduce security obligations if the output is stored unredacted. Define retention, access and redaction before sending the audio. For healthcare deployments, the official documentation flags the need to speak with Sales about a Business Associate Agreement where HIPAA compliance is required. Do not infer compliance from the existence of a medical model alone.
Speaker and timing QA
Have a reviewer check speaker switches, overlap, quiet utterances and proper names. Use word timestamps as starting points for subtitles, then correct reading speed and line breaks. If speakers shout over each other, diarization can produce plausible but wrong turns. Preserve an export that contains model ID, original filename hash and all manual corrections; never overwrite the unedited recognition trace.
Alternatives
AssemblyAI’s Universal-2 has broad language documentation; Deepgram Nova-3 multilingual supplies language-hint controls; Whisper offers local execution. Compare vendor claims under one independent corpus including accented and code-switched speech. Price by recording duration plus any supplementary features in the external pricing ledger.
Worked production example
A podcast may contain two presenters, a remote guest, opening music and several overlapping jokes. Diarization can mark the voices, yet the labels might swap after a musical interlude. Ask an editor to mark speakers on a small gold-reference sample and check each switch in the machine output. Speaker 0 and Speaker 1 are arbitrary labels; attach real names only after confirming the guest list and listening to the segment. Keep a separate record for inaudible or uncertain speech rather than inventing an authoritative quotation.
Practical acceptance gate
In a caption pipeline, use word timestamps as raw timing, merge them into legible cues and validate against audio at ordinary speed. For a live event, don’t publish partial hypotheses to permanent archives until a final pass is available. Test whether entity detection or prompts alter important private information and keep the full transcript behind appropriate access controls. Choose batch, realtime or medical model variants by the documented purpose; do not assume identical capabilities simply because they share the Scribe brand.
Region, plan and model-version notes
Documentation checked on 2026-10-25. Exact applicability: availability varies by account/endpoint and jurisdiction. Version: scribe_v2 / scribe_v2_realtime. Plan: not independently verified for a specific plan. Monetary prices and quotas are not asserted unless separately verified; check the canonical AI 10 pricing ledger before purchase or publication of a tariff.
Sources & Further Reading
- Eleven Stt — official documentation (checked 2026-10-25)
- Assembly — official documentation (checked 2026-10-25)
- Deep Nova — official documentation (checked 2026-10-25)