Zoom Scribe leads a final-transcript test: what to check

Artificial Analysis has placed Zoom Scribe V2.1 Pro Live first in its published AA-WER Streaming final-transcript comparison, reporting a 2.4% word error rate against 2.5% for…

JJames Bennett1 min read

Artificial Analysis has placed Zoom Scribe V2.1 Pro Live first in its published AA-WER Streaming final-transcript comparison, reporting a 2.4% word error rate against 2.5% for MAI-Transcribe-2-Streaming. The evaluator shared the result on 19 November.

For teams evaluating speech recognition for customer-support tools, this is a useful shortlist signal. It is not a measurement of how accurately a complete support agent answers a customer, handles an order or resolves a return.

Read the metric before the ranking

The streaming leaderboard distinguishes the final transcript from the first partial transcript and measures their timing separately. Its roughly eight-hour audio evaluation combines AA-AgentTalk, VoxPopuli and Earnings22, weighted 50%, 25% and 25% respectively.

A low final-transcript error rate therefore answers a different question from whether an agent receives usable text quickly enough to respond naturally. The ranking also depends on the benchmark’s audio and scoring method; it does not establish the error rate on your customers’ calls.

Test the commerce details that matter

Our recommended evaluation includes product names, order references, addresses, accents, interruptions and noisy calls drawn from appropriately authorised test data. Score important entities separately from overall transcript quality: one wrong product code can matter more operationally than several harmless wording differences.

Measure response timing and downstream task success alongside transcription, and compare the full service cost at your expected volume. Keep a human escalation path for uncertain order details rather than allowing a plausible transcript to become an unchecked action.

Zoom’s Live-mode announcement dates to 31 July. This article covers the newly published benchmark result, not a claim that live transcription itself launched this week. No merchant deployment outcome is established by this test.

ℹ️
Final-transcript accuracy is one measure. Evaluate partial results, response timing and order-detail accuracy before choosing a support stack.

Read the announcement if the embed does not load.

Filed underAI

More from AI

See all →