How it works
Every app passes through the same two stages. Gates remove apps that cannot work for you at all; scores order the survivors by fit for your use case, language and privacy stance.
Two stages: gates, then weights
The ranking has two stages that do fundamentally different jobs. Gates answer the question 'could this app work for this person at all?' Scores answer 'of the apps that could work, which fits best?' Keeping them separate matters: an app should never rank highly for someone whose hard requirement it fails, no matter how good it is otherwise.
This is the difference between filtering and ranking. A Windows-only user should never see a Mac-only app at position seven 'with caveats' — it should simply not be in the list. Gates do that. Everything that survives the gates is then scored on the same scale.
Stage one: the hard gates
Platform is the simplest gate. If you say you work on Windows, apps that only ship for macOS leave the list. This sounds obvious, but most 'best dictation app' articles on the web don't do it — they rank Mac-only apps for everyone, which is how Windows users end up three paragraphs into a review of software they cannot install.
Privacy stance is the second gate, and it is a gate rather than a score because it behaves like one. If you require fully on-device processing — because of employer policy, client confidentiality, regulation, or principle — then a cloud-only app is not a worse option, it is a non-option. Choosing 'fully on-device' removes every app that ships your audio to a server, even if that app tops every other measure. Choosing 'hybrid' keeps apps that can run locally even if they also offer cloud features.
Stage two: the scoring formula
Each surviving app carries a set of scores from 0 to 10: a per-language dictation accuracy score for the language you picked, and a fit score for each of the five use cases. The ranking multiplies each score by its weight and sums the result. Your language choice controls which accuracy score is used; your use case controls which fit score dominates the weights.
Language accuracy is the heaviest single factor, because an app that mishears you fails at its one job regardless of features. Within it we weigh punctuation and formatting quality for your language, not just raw word error rate — the difference covered in our language-support guide.
Price enters as a modest negative weight rather than a gate, because budgets differ too much for a hard rule. A free open-source app and a $30/month app can both legitimately top the list for different people with the same language and use case.
What your use case changes
Each use case re-weights the same underlying scores. Coding privileges identifier handling, command modes and latency. Long-form writing privileges formatting quality, paragraph structure and tolerance of thinking-out-loud pauses. Email and chat privileges speed and automatic tone cleanup. Clinical and legal work privileges vocabulary accuracy, on-device options and compliance posture. Accessibility privileges reliability, hands-free control and error recovery.
Nothing about the app changes when you switch use case — only the weights do. Wispr Flow is the same product for a novelist and a developer; it ranks differently for them because the two jobs reward different capabilities.
A worked example
Suppose you dictate in German, your use case is long-form writing, you require fully on-device processing, and you work on a Mac. The platform gate removes the Windows-only apps (SpeechPulse, Dragon Professional). The privacy gate removes every cloud-only app (Wispr Flow, Aqua Voice, Willow Voice and friends).
The survivors — Superwhisper, MacWhisper, VoiceInk, Handy and the other local apps — are then scored. VoiceInk gets a small edge from its Parakeet V3 engine option, since German is one of the 25 languages Parakeet handles better than Whisper. Superwhisper scores well on formatting polish. Handy scores well on price — it is free — and loses a little on polish. The order you see is that arithmetic, nothing else.
What the model deliberately does not do
We do not accept payment for placement, and no app's score changes because of an affiliate relationship. Where a link earns us a commission it is labelled, and the ranking code has no access to which links those are.
We do not average in user reviews from app stores. They measure satisfaction with the whole product — price, support, onboarding — rather than dictation quality, and they are too easy to inflate. Our accuracy scores come from our own test corpus, dictated identically into every app.
We do not score features that don't affect dictation output. A beautiful settings window is nice; it does not move the number.
Frequently asked questions
Why did an app disappear from my list entirely?
It failed a hard gate — almost always platform (it doesn't run on your OS) or privacy (you required on-device processing and it is cloud-only). Change either input and it will reappear.
How often do scores change?
We re-run the dictation test corpus when an app changes its engine or a major version ships, and we note the update date on each review. App updates that only touch UI don't move scores.