Artificial Analysis: Zoom’s streaming model most accurate on final transcripts, with a 2.4 per cent word error rate
In Artificial Analysis’s AA-WER Streaming test, Zoom Scribe V2.1 Pro Live had a final-transcript word error rate of 2.4 per cent, narrowly below MAI-Transcribe-2-Streaming’s 2.5 per cent. It costs $0.40 per hour of audio.
Artificial Analysis, which compares AI models, said on its X account on 19 November that it had tested Zoom’s new Zoom Scribe V2.1 Pro Live model itself. The model converts speech to text as it is spoken.
The metric is word error rate (WER), the share of words transcribed incorrectly. According to the image in the post, the AA-WER Streaming test counts errors in the final text after the system detects that speech has ended. The figure therefore does not show how accurate the interim text displayed during speech is.
The gap between the top two results is small: Zoom’s 2.4 per cent and MAI-Transcribe-2-Streaming’s 2.5 per cent differ by 0.1 percentage points in a single test. The gap does not show that Zoom’s model would transcribe other kinds of audio clearly better than its rival.
Artificial Analysis said Zoom’s model costs $0.40 per hour of audio, or $6.67 per 1,000 minutes. It is the Pro tier of Live mode in the Zoom Scribe API. Per 1,000 minutes, MAI-Transcribe-2-Streaming costs $9 and ElevenLabs’ Scribe v3 Live $3.67.
The models differ more on price than on error rate. The image in Artificial Analysis’s first post gives ElevenLabs’ model a word error rate of 2.9 per cent in the same test. Because a small error gap in one test does not show how the models handle other audio, it may be worth trying the cheaper model on your own recordings.