Model releases
A better captioning tool can still get the name wrong
A short recording with two speakers can reveal which names, times and corrections ElevenLabs’ new Scribe v3 misses.

A caption can miss a name and still look ready to publish. ElevenLabs announced Scribe v3 for recorded audio on 10 November and updated the announcement on 13 November. It also offers Scribe v3 Live for speech as it happens. A Caribbean newsroom, training team or event producer can try the recorded version on a short clip before moving a whole archive. Check the name against the recording, then replay any correction to a time or amount.
A test with brand names exposed a weakness
In ElevenLabs’ own test, Scribe v3’s word error rate fell from 16.94% to 12.38% across 102 languages, a relative drop of about 27%. Word error rate compares a transcript with a reference and counts missing, added and substituted words. The test used FLEURS, a collection of short clips of people reading aloud, with one speaker per clip. Artificial Analysis put the recorded Scribe v3 at 1.9% and Scribe v2 at 2.2% on its own index, checked on 14 November. Its test weights recordings of people speaking to voice agents, public debate and company earnings calls. The scores come from different recordings and methods. Neither test tells you how Scribe will handle a busy room in Kingston.
The Sonetho team published a small comparison on 11 November. It used two ElevenLabs voices to read seven Korean scripts, then sent the files to Scribe v2 and v3 with automatic language detection. On a clip heavy with brand names, the new version scored about 60% character errors against 25% for v2. It sometimes wrote the names phonetically in the Korean alphabet. That can preserve the pronunciation while losing the spelling an editor needs. Both models missed Kling 4.0. Sonetho counted changes to individual characters and ignored spaces and punctuation. That is one clip made with synthetic voices. Use the result to decide which details to test on your own recordings. Tellers, an existing video editor, says it added the recorded model for captions on 11 November. Its report gives no measured change in editing time.
Give two speakers one short clip
For a practice recording, ask two colleagues to read these lines into a phone, with a brief overlap on the reply. Speaker A: “Meet Rochelle at the Harbour View desk on Tuesday at 3:15. The group has fifteen seats.” Speaker B: “No, the Friday group has fifty seats. Tuesday still has fifteen.” Upload the same clip to ElevenLabs’ recorded-audio tool and to your usual service. ElevenLabs lists limited free transcription and requires an account. Its published restrictions include Cuba. Jamaica is not listed as blocked, but access was not tested from a Jamaican connection. Use a practice clip or material your team has permission to upload.
Keep both raw transcripts. Give the person reviewing them this instruction: “Play the clip and flag any mistake in Rochelle, Harbour View, Tuesday at 3:15, Tuesday’s fifteen seats, Friday’s fifty seats or which speaker corrected the count. List each fix. Leave the rest of the wording alone.” Time the full job through review and export. If both transcripts can be fixed to match the clip, choose the service with the lower total review burden. Replay the final captions on a phone before publishing. A corrected transcript can still have captions that arrive too soon or disappear too quickly.
Ask the seller who carries the risk
Use the AI Vendor Assessment when a seller offers to take over captioning. Select Score This Vendor, work through the questions for one proposal and choose View Results Without Sharing Contact Details. For the contract questions, try a proposal with a paid pilot and loosely set success criteria, a difficult exit and no remedy if the captions miss the agreed standard. Those choices produce a Commercial Protection score of 3 out of 9. The score comes from your answers. You still need to test the captions and read what the contract promises if they fail. Ask the seller to put export rights and a pass condition for names, amounts and speaker changes in writing before you pay for a larger job.
Sources and further reading
- ElevenLabs’ Scribe v3 and v3 Live announcement
- ElevenLabs model documentation
- Artificial Analysis speech to text comparison
- Sonetho’s Korean audio test
- Tellers’ Scribe v3 caption workflow
- ElevenLabs’ transcription plans
- ElevenLabs’ country restrictions
- AI Vendor Assessment
Sources checked 14 November 2026.