ElevenLabs Scribe v3 claims 27% fewer word errors, live model 43%, at unchanged prices
ElevenLabs has upgraded its speech-to-text line. Scribe v3 reduces word errors by 27% versus its predecessor, and Scribe v3 Live, the real-time streaming variant, reduces them by 43%. Prices stay the same, so existing customers get the accuracy gain at no extra cost. An independent evaluation on phone-call audio, a hard domain with compression, crosstalk and background noise, placed Scribe v3 Live first in two of the four languages tested.
The larger improvement in the live model is the more commercially significant number. Streaming transcription feeds voice agents, call-center analytics and live captioning, where errors propagate immediately into downstream reasoning. A voice agent that mishears an account number or a drug name fails in front of the customer. Batch transcription can be corrected afterwards; live transcription cannot. Cutting live errors by nearly half narrows the gap between streaming and offline quality, which has long been the main trade-off in the category.
Competition in speech recognition has intensified. Mistral highlighted Voxtral Transcribe 3 this week, using context biasing with speaker names and technical terms to transcribe every main-stage talk at AI Engineer Paris in real time. OpenAI is moving ChatGPT Voice users entirely to GPT-Live and retiring its legacy voice modes on December 11. Deepgram, AssemblyAI and the hyperscalers all compete on word error rate per dollar. ElevenLabs' position is unusual because it sells both synthesis and recognition, so it can offer a full voice-agent loop from a single vendor.
Caveats: the 27% and 43% figures come from ElevenLabs, and the baselines and test sets were not detailed in the coverage. The third-party phone-call test is encouraging but covered only four languages, and Scribe lost in the other two. Teams evaluating it should run their own audio, especially accented speech, domain jargon and noisy channels, before switching. Watch whether competitors respond with price cuts, since holding prices while improving accuracy puts pressure on the rest of the market.