AI Aulis

AI NEWS IN ENGLISH · 22 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

21 stories · next issue tomorrow morning · How the paper is made →

← Issue 22 November 2026Paper

Coval benchmark: Zoom’s Scribe Pro v2.1 has the lowest word error rate of 31 models, at 1.0 per cent

AI-generated contentWritten by AI

benchmarks.coval.ai · Source publication date unknown

Zoom’s Scribe Pro v2.1 is the most accurate model in Coval’s speech-to-text benchmark, but it ranks only 22nd for latency. The fastest model, Saidwire’s Wire-1, has a p95 latency of 34 milliseconds: in 95 per cent of measurements, the final transcript came back within that time.

Coval compares speech-to-text models on two metrics. One measures how many errors a model makes, the other how long a voice agent waits for the text. The latest data on Coval’s page is from 03:00 GMT on 22 November, and the tables cover the last 24 hours.

Word error rate (WER) shows the share of words a model gets wrong, leaves out or adds. Coval breaks Scribe Pro v2.1’s 1.0 per cent down into substitutions (0.6 per cent), deletions (0.2 per cent) and insertions (0.2 per cent). The figure is pooled across all datasets in the benchmark. Some datasets are WildASR variants, in which clean audio has been degraded in one way.

Coval measures speed with TTFS (time to final segment): the time from the end of speech until the final transcript is returned. This is how long a voice agent waits before responding. The figure excludes the time the system spends detecting that the turn has ended. Coval’s p95 figure means that 95 per cent of measurements were faster than that.

The table shows p95 latencies for selected models. They reveal nothing about accuracy: the most accurate model ranks 22nd for latency.

Model Provider p95 TTFS
Wire-1 Saidwire 34 ms
Rumble-2 Reson8 48 ms
Scribe v3 Live ElevenLabs 107 ms
Nova 3 Deepgram 134 ms
Universal 3.5 Pro AssemblyAI 146 ms
Scribe Pro v2.1 Zoom 356 ms
GPT Realtime Whisper OpenAI 817 ms
Chirp 3 Google 959 ms
Gemini 3.5 Transcribe Live Gemini 1,747 ms

Source: Coval, 24-hour measurement period, data as of 03:00 GMT on 22 November 2026. Scribe Pro v2.1 has 584 samples; most other models have 912–916.

To help users choose, Coval plots a Pareto frontier. Coval says it shows the best accuracy available at each latency budget. A model outside the frontier is beaten on both speed and accuracy by a model on the frontier.

The text version of the page does not show the other models’ word error rates, so their accuracy ranking cannot be read from it.

What this is based on

  1. Speech-to-Text Benchmarks

    Publisher: benchmarks.coval.ai · source publication time unknown · retrieved 22 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

Previous coverage

Report an error in this story