Zoom Scribe V2.1 Pro Live tops streaming WER but trails on latency
- Source
- x.com
- Date
Zoom Scribe V2.1 Pro Live takes the #1 spot for Final Transcript accuracy on AA-WER Streaming, with a 2.4% word error rate, just ahead of MAI-Transcribe-2-Streaming at 2.5% We benchmarked @Zoom 's new streaming Speech to Text model independently. Congratulations to the Zoom team on the launch! 🧵

Context
Zoom's new live transcription model, Scribe V2.1 Pro Live, makes the fewest errors in its finished transcripts of any model on the streaming leaderboard kept by Artificial Analysis, an independent benchmarking firm, but it takes longer than its closest rivals to finish. transcription works on audio as it arrives, as live captions or a voice assistant need, rather than on a finished recording. Such a model sends provisional text, called partials, that can still change while someone speaks, and then confirmed text, called finals. Artificial Analysis detects when the speaker stops, requests the final transcript where the model allows it, and scores that transcript by word error rate, the share of words swapped, dropped or added compared with a human-checked reference.
On that measure Zoom's model scores 2.4%, just ahead of Microsoft AI's MAI-Transcribe-2-Streaming at 2.5%, Artificial Analysis reports. The leaderboard data shows what the lead costs. Zoom's final transcript arrives about 0.66 seconds after the speaker stops, against 0.13 seconds for Microsoft's model and 0.08 seconds for ElevenLabs' Scribe v3 Live, which scores 2.9%. Zoom's first provisional text after the speaker stops is also much rougher, at 5.0% errors, while Microsoft's is already at 2.5%.
Which model fits depends on what waits for the transcript. Artificial Analysis built this for voice agents, and its launch article explains that a faster transcript leaves more of the response time for the to reason and reply. For an agent answering a caller, Microsoft's model gives nearly the same accuracy five times sooner, though Artificial Analysis lists it at $9 per 1,000 minutes of audio against $6.67 for Zoom's. For meeting notes or transcripts read afterwards, where nothing has to reply, the extra half second matters less and Zoom's model gives the most accurate settled text on the board.
- streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Checking sign-in…
Loading comments…





