Dictation glossary

Text cleanup

Text cleanup means editing a raw speech-to-text transcript so it looks like written language: punctuation, capitalization, filler-word removal, and user-defined replacements. Rule-based cleanup uses lists and patterns. Optional language-model cleanup rewrites the sentence more freely while trying to keep your meaning. It happens after recognition, on text, not on the audio buffer.

Get WhisperJot Try the browser demo
In more detail

What is Text cleanup?

A recognizer emits words it heard, often without reliable commas or paragraphs. Rule-based cleanup is deterministic: strip um, expand a spoken command, fix spacing around punctuation, apply a replacement table. You can read the rules and predict the output. That predictability is why rules are the right default for code identifiers and legal phrases you must not paraphrase. They fail on messy grammar that is still what you meant.

A local language model, reached at 127.0.0.1 if it stays on the machine, can punctuate and tighten prose that rules cannot. It is extra RAM and an extra failure mode: over-editing, delayed insert, or a rewrite that sounds unlike you. Cloud rewrite APIs add a second network hop for text that may already be sensitive. The honest stack is layered: recognition, then rules you control, then an optional model you opted into, running locally if privacy is the point of the app.

For writers who speak

Why it matters for dictation

Raw ASR reads like a court reporter who never went to school. Cleanup is the difference between dictation you send and dictation you rewrite. If cleanup is only a remote prompt, you bought a second cloud. If it is only a toy regex, long sentences stay broken. Writers should know which layer they are trusting.

In this product

How WhisperJot handles it

WhisperJot applies rule-based cleanup — fillers, punctuation, vocabulary, replacements, and voice commands such as new paragraph — after every dictation. Optional LLM polish runs locally on 127.0.0.1 via MLX-LM or Ollama. Transcription itself stays on-device by default; Jot Cloud is opt-in and still goes through the same cleanup stack.

Related terms

Keep reading in the glossary.

Filler-word removal

Filler-word removal means deleting spoken hedges — um, uh, like, you know — from a transcript so the remaining text reads more like writing than like a raw recording.

Read Filler-word removal →

Custom vocabulary

Custom vocabulary means a user-supplied list of words, names, and spellings that a speech-to-text system should prefer when it hears something similar.

Read Custom vocabulary →

Text injection

Text injection means placing transcribed words into the application that currently has keyboard focus.

Read Text injection →

Live partial transcription

Live partial transcription means showing in-progress text while audio is still being captured, instead of waiting until you stop talking to reveal a finished transcript.

Read Live partials →
Questions

Straight answers.

See also: Pricing, System requirements.

What is text cleanup in speech-to-text?

Text cleanup is everything that happens to a transcript after words are recognized: punctuation, capitalization, filler removal, replacements, and sometimes a language-model rewrite. It does not re-hear the audio. The goal is text you would actually paste into a document. Rules are predictable; models are flexible. Most dictation apps combine at least a light rule pass with the recognizer's own decoding.

What is the difference between rule-based cleanup and an LLM pass?

Rules match patterns you can list: delete um, capitalize I, replace a spoken shortcut. An LLM pass reads the whole sentence and may rephrase. Rules rarely invent facts; they can miss messy speech. Language models can over-smooth and need to stay on 127.0.0.1 if you want the polish to match an on-device privacy story. Use rules first; add a model if the residue still looks spoken.

Does text cleanup run in the cloud?

It can, but it does not have to. Rule tables are just data on disk. A language-model pass can call a remote API or a local server. If the rest of the app is on-device, sending the transcript out for 'polish' reintroduces a network copy of your words. Ask where that step runs, and whether it is off unless you enable it.

One hotkey, any focused app.

Private local transcription by default, with an optional opt-in cloud engine.

Get WhisperJot — $99/year