Wispr Flow ranks #1 on Voice Arena's independent Dictation Bench
Four months ago, we launched a lab. Two months ago, we launched our first model. Now, Wispr Flow is #1 on Voice Arena's Dictation Bench.
Voice Arena's independent Dictation Benchmark tested 10 Mac dictation apps with multiple challenges seeking to answer one question: could you send the output without editing it? Flow came out on top, with 79.4% of its drafts ready to send as is.
The lead was even wider on how much editing a draft needed. Voice Arena counted the characters a person would have to change, a measure called AMEND, and Flow needed less than half the editing of any other app: 0.40% of characters, against 0.93% for the next-best.
Accurate words got us partway there. Getting to ready-to-send took our Canto model and our product working together.
Why we built a lab
As more people talk to their devices to do real work, we set out to make dictation you never have to fix.
Dictation isn't transcription. When you dictate, you're not reading from a script. You say a colleague's name, a product, a term from your field. You start a sentence, change your mind, and start again. And you want the text the moment you stop talking.
Each of those moments can end the same way: someone stops to fix what they dictated. After decades of typing, people only hand their writing to voice once they trust what comes back, and every fix teaches them that they still need to check.
We launched the Wispr Advanced Interfaces Lab in July and trained our own model, Canto 1. We built it for real dictation, not clean recordings: phones on a counter, windy streets, laptop mics, earbuds, background voices, traffic, whispers, and two-word replies. It learns from people's corrections, separating true recognition errors from ordinary rewrites, and still answers in real time.
Voice Arena's Dictation Bench asks the same question we built Canto around: could you send what you said without touching it?
Why Flow needs so few edits
We've always judged Flow by Dictation Edit Rate: how often people edit a dictation after it appears. So we trained Canto on the edits people actually make. It separates true recognition errors from ordinary rewrites and learns only from the errors. It also uses context selectively: when a name in your dictionary sounds close to what you said, it weighs that against what the audio supports.
The result is a model that rarely makes a big mistake. When Flow was wrong in this test, no draft needed more than 19 characters of fixing. Voice only saves time if you can rely on what comes back, and once people stop checking, they hand over more of their typing.
What's next
Voice Arena tested Flow on Canto-1. Our next model, Canto-2, launches soon, and we'll judge it the way we judge everything we ship: by how often people have to fix what Flow writes. Every percentage point matters. Across millions of dictations, each point means thousands more messages sent without reaching for the keyboard.
View the Dictation Bench leaderboard and read the full white paper at Voice Arena. You can learn more about the Wispr Advanced Interfaces Lab here.
