# Rohan Paul calls for AI dictation apps to be judged on Dictation Bench

> Original title: If a dictation company says it is the best, I would now ask how it scores on Dictation Bench.

- Source: Rohan Paul
- Published: 2026-11-11T15:09:54.000Z
- Bad Influence: https://badinfluence.news/items/kjt18zc0mzxluqs3k7q0ndg0i
- Original: https://x.com/rohanpaul_ai/status/2120534491194364310

## Summary

Rohan Paul says a dictation company's "best" claim should be checked against Dictation Bench, a shared test from Voice Arena. The benchmark plays the same 761 recordings from 87 speakers into ten Mac dictation apps and publishes its policies, configurations, judge instructions and results. Paul says the current US English, macOS-only scope should be kept in mind, and he wants companies to report against the benchmark routinely.

## Post

If a dictation company says it is the best, I would now ask how it scores on Dictation Bench. 
One test does not cover every use case, but that kind of claim should hold up in a shared comparison that anyone can inspect.

@voicearena_ai played the same 761 recordings from 87 speakers into ten dictation apps on a Mac (the real apps, not their APIs). They published the reference policies, the app configurations, the judge instructions, and results that flag which gaps are statistical ties.

So companies now have something concrete to respond to. They can show where their product leads, explain the failures, or show that a different task changes the result. Any of those is more useful than asking users to compare percentages measured under different conditions.

The current test covers US English on macOS only, and the results should be read with that scope. Within it, I would like companies to report against this benchmark routinely. A shared standard makes claims easier to check and improvements easier to prove, and this benchmark is a practical place to start.

## Quoted post @shobhitbanga

> Drop 4/15: Announcing the Voice Arena Dictation Bench - the first open benchmark of AI dictation apps.
> 
> Not their APIs: the real apps. We played 761 recordings from 87 speakers into each of 10 apps on a Mac, then 5 LLM judges asked if each draft could be sent as-is.
> 
> Wispr Flow: 79%. Apple: 35% 🧵

https://x.com/shobhitbanga/status/2120527865947234622