AI WORLD HQ
Think about a meeting where two people talk over each other, then read the notes the next day and find a decision credited to the wrong person. That small mistake is what speech-to-text (software that turns spoken audio into written words) is trying to prevent.
Why transcripts go wrong
ElevenLabs points out that real-world audio is rarely tidy. People interrupt each other, pause in the middle of a sentence, or switch languages while they talk. The company says a mistake early in a transcript travels into everything built on top of it, such as summaries, subtitles, and follow-up notes, and those errors cost more to fix the later you find them.
What's new in Scribe v3
ElevenLabs describes two models with different jobs:
- Scribe v3 is for recordings you process after the fact. The company says it makes 27% fewer word errors than Scribe v2 across 102 languages. Word error rate (WER) is the share of words a system gets wrong, so a lower number means a more accurate transcript.
- Scribe v3 Live is the low-latency version (latency is the short delay between someone speaking and the text appearing), built for live audio. ElevenLabs says it makes 43% fewer errors than Scribe v2 Realtime.
Speaker labels are a major focus. Diarization (working out who said what) is now part of the real-time model for the first time. The company says Scribe v3 makes 17% fewer diarization errors than its predecessor on recordings with three or more speakers, which is where many tools struggle. Scribe v3 Live also brings diarization to realtime transcription for the first time, with 39% fewer diarization errors than the next-best realtime model tested.
Turn detection (deciding when someone has finished speaking) is another change. ElevenLabs says Live looks at the context of the conversation as well as silence. That means a caller who pauses halfway through a phone number is less likely to get cut off.
You also get more control over the output:
- Choose whether numbers appear as digits or as words.
- Give the model up to 1,000 key terms, such as brand names, acronyms, or jargon, so it spells them correctly.
- Write formatting rules in plain English, such as asking for dates in a consistent order.
- Use a multilingual mode in Live for conversations that switch languages mid-sentence.
- Use a keyterm library for much larger vocabularies, like product catalogs. The model checks it only when it hears a likely match.
How the numbers were measured
ElevenLabs tested both models against its earlier versions and other popular speech-to-text models, using the same datasets for each. Those results are the company's own claims, and the company says it will open source its real-world test dataset soon. Treat the percentages as the vendor's measurements, not independent results.
Wrap-up
Scribe v3 and Scribe v3 Live are ElevenLabs' newest speech-to-text models. They focus on fewer errors, better speaker labels, smarter pauses in live conversations, and more control over names and formatting. The company's numbers are its own, but the features are clear enough to check for yourself. For your next step, pick a recording you already have, listen while you read the transcript, and note how it handles names and who said what. For pricing, check the official ElevenLabs pricing page.