
The results are in: Aqua Voice is the most accurate dictation tool in the world. Voice Arena, an independent benchmark, tested fourteen voice input tools and ranked Aqua Voice first.
This benchmark is the first to move beyond the Word Error Rate (WER) metric traditionally used to evaluate transcription services. The test instead measures how often a dictation could be sent with zero edits. Aqua passes this test 80.5% of the time, more than any other app. On the hardest subset of dictations in the evaluation, the gap between Aqua and second-best was even wider.

Wispr Flow, who makes a similar app, came in second behind Aqua on the benchmark. Willow Voice, inventors of an animated character called Wally, came in fourth.
Full Results
Voice Arena measures how many drafts need zero edits before a user would send them: the Zero Edit Rate (ZER). The idea is to identify the precise characters the user would have typed and check whether the dictation output is an exact match. Aqua Voice was the best performer of the tested systems on this measure.

We attribute this score to Avalon, our family of in-house voice models. They are trained on many audio tasks for maximum accuracy on real-world voice input. The table also lists AMEND, the share of characters you would have to fix across all of an app's drafts, where lower is better.
Rank | App | Zero Edit Rate | AMEND |
|---|---|---|---|
1 | Aqua Voice | 80.5% | 0.54% |
2 | Wispr Flow | 79.3% | 0.40% |
3 | Blurt | 78.0% | 0.60% |
4 | Willow Voice | 76.5% | 0.65% |
5 | Gemini 3.5 Transcribe | 75.8% | 1.23% |
6 | Spokenly | 73.8% | 0.64% |
7 | Granola | 71.9% | 1.03% |
8 | Kivi | 69.3% | 1.34% |
8 | Meta AI | 69.3% | 2.56% |
10 | ChatGPT Dictation | 64.3% | 1.15% |
11 | Claude Dictation | 64.0% | 0.92% |
12 | Superwhisper | 59.8% | 1.45% |
13 | Perplexity Dictation | 52.7% | 1.85% |
14 | Apple Dictation | 34.5% | 2.60% |
How it was tested
Voice Arena recorded 87 people dictating the things people use dictation for: work chat, email, personal messages, prompts for AI tools, code, notes and longer drafts. It played the same 759 recordings into every app on a Mac, with each app at its default settings, and saved what each one typed. The test covers US English.
The outputs were judged by five different LLMs, each making the fewest number of edits needed to turn the draft into what the speaker meant to write (intent). If no edits are required by at least 3 of 5 judges, it is marked perfect.
What about Word Error Rate?
Most dictation benchmarks use Word Error Rate (WER). We've written previously about why that's not the right way to measure. WER measures typically:
- ignore differences in punctuation and capitalization
- are inconsistent in their handling of false starts, partial words, and user self-corrections
- count minor errors that don't change the meaning the same as major errors, like an incorrect name or a dropped "not"
Voice Arena is one of the first third parties to adopt an approach that avoids these problems. It would be great to see others like Artificial Analysis and OpenASR follow suit.
How Aqua Voice compares to other apps
Aqua Voice targets the Pareto optimum of accuracy (what the user meant) and speed (latency). These results show that despite a median latency of 365 ms, accuracy is not just good, but state-of-the-art.
On the same test, Willow Voice scored 76.5%, Superwhisper 59.8% and Apple Dictation, the dictation built into macOS, 34.5%. 9to5Mac saw a similar gap when it read the same speech into both: Apple Dictation made 17 mistakes, and Aqua Voice made one.

For features, platforms and pricing, see how Aqua Voice compares with Wispr Flow, Willow Voice and Superwhisper.
Check the results yourself
Voice Arena publishes its methodology and the full leaderboard, and its test harness is open source.
If you haven't tried Aqua Voice in a while, download it and dictate the next thing you would have typed.


