Available in Ollang Workflows as
elevenlabs-scribe-v2 — Closed Captions, Subtitle Translation, and AI Dubbing. See the Speech-to-Text catalog.Models
Published accuracy tiers
Rather than quoting a single headline accuracy figure, ElevenLabs classifies each supported language into a word error rate band, from Excellent (5% WER or below) down to Moderate (25–50% WER). This is the single most useful thing about the provider from a workflow-planning perspective: you can check the tier for your specific language before committing a Workflow, instead of discovering the answer in review. Treat the language list as a set of tiers, not as a flat 90.Key capabilities
- Speaker diarization up to 32 speakers — the highest documented ceiling in the catalog, which matters for panels, roundtables, town halls, and large meetings where other providers collapse speakers together.
- Dynamic audio tagging — non-speech events such as laughter and applause are captured as
audio_evententries rather than dropped, which feeds accessibility caption requirements directly. - Entity detection across 65 types, for structured extraction alongside the transcript.
- Keyterm prompting with up to 1,000 terms — a large vocabulary budget for enforcing product names, character names, and domain terminology.
- Precise word-level timestamps, giving subtitle segmentation more to work with, which shows up in reading-rate compliance.
Where it fits in a workflow
Reach for Scribe v2 when the source is in a language outside the mainstream 30, when the recording has many speakers, or when non-speech audio events need to appear in the captions. For monolingual, mainstream-language content it and AssemblyAI are close substitutes, so the choice comes down to which one benchmarks better on your material — and AssemblyAI’s Universal-3.5 Pro tier is the stronger documented option for English–Spanish code-switching specifically.Trade-offs
- Accuracy varies widely by language. The published tiers span under 5% WER to over 25%; a language in the lower bands will need review time budgeted accordingly.
- On unclear or silent passages, transcription models of this class can emit fluent text that was never spoken. That failure mode does not look like an error in the transcript, so validate anything going straight to a deliverable.