How it works
Two AI systems, not one: the speech model turns audio into raw words, then a language model turns raw words into the text you meant to write.
Why the cleanup pass exists
Nobody speaks the way they write. Spoken language is full of fillers, false starts, repetitions and half-finished sentences that get abandoned mid-thought. Classic dictation software transcribed all of it faithfully, which is why classic dictation produced text that needed heavy editing and why most people who tried it gave up.
The insight that built the current generation of dictation apps — Wispr Flow, Aqua Voice, Willow Voice and the rest — is that transcription was never the hard part anymore. Whisper-class models solved recognition in 2022. The hard part is that faithful transcription of real speech is not usable text. So these apps run a second model after the first: a language model that reads the raw transcript and produces the sentence you were trying to say.
What the layer actually does
Four jobs, in order of reliability. Filler removal: 'um', 'like', false starts and self-corrections are stripped. This works nearly perfectly. Punctuation and capitalisation: spoken language carries punctuation in rhythm and pitch rather than explicit marks, and the model reconstructs it. This works well for major languages and less well for smaller ones.
Formatting: 'new paragraph', lists, dates and numbers converted to written form. Also reliable, and the feature that saves the most time. And finally tone adjustment: some apps detect whether the target application is Slack, Gmail or a code editor and adjust register accordingly — casual in chat, formal in email. This is the newest and least predictable of the four.
What it costs
Latency. A second model pass adds a delay between when you stop speaking and when text lands. The best implementations hide it with streaming; the worst make you wait a beat after every sentence. Whether this bothers you depends on whether you dictate in bursts or continuously.
It also costs money, which is why the feature is concentrated in subscription apps: every dictated sentence costs a model call. The local apps are catching up — small language models can do filler removal on-device now — but the most aggressive cleanup still happens in the cloud, which is one more reason the privacy question in this category can't be dodged.
Where it goes wrong
The failure mode to know about is confident rewriting. A cleanup model that isn't sure what you said will sometimes produce a fluent, grammatical sentence that is not what you meant — a changed number, a softened claim, a name swapped for a similar one. Because the output reads so cleanly, these errors are harder to catch than classic transcription gibberish.
Practical habits that protect you: read numbers and names back before sending anything consequential; learn your app's verbatim mode (most cleanup apps have one) for text where exact wording matters; and be suspicious of any sentence the app produced that reads better than you think you speak.
Frequently asked questions
Can I turn the cleanup layer off?
Most modern apps offer a verbatim or literal mode that bypasses the rewrite and gives you raw transcription. It's usually a per-app or per-hotkey setting — worth finding before you dictate anything legal or technical.
Does the cleanup pass see my audio?
No — it sees the text transcript. But in cloud apps both steps happen on the vendor's servers, so the privacy question applies to the pipeline as a whole, not to either model separately.