Dictation Cleanup Changed My Meaning: What to Check

To remove “um” and repeated starts without reorganizing your message, start with Settings → Cleanup → Standard. If exact wording matters, choose Verbatim (punctuation only). If the result changes a number, a “don’t,” or a “maybe,” compare raw and final text in Dictation History before sending it.

Removing hesitation and rewriting a request are different jobs. A shorter sentence is useful only if the same person, action, conditions and uncertainty remain. This maker-authored guide explains Ducduc’s controls and limits; one example preserves an actual text-only backend response; other examples are marked illustrative.

First recover the wording you need

1. Open Dictation History and find the affected entry. Show the raw transcript when it is available alongside the final result.
2. Copy the wording you trust into the destination and correct it before sending. Raw text can contain recognition errors too.
3. Find the first difference that changes meaning. Record the small raw/final pair rather than describing the whole result as “bad.”
4. Change one setting and test another short, non-sensitive sentence in Notes before returning to your work.

What you find	Check first
The raw transcript already has the wrong number or word	Recognition language, microphone and recurring vocabulary
Raw text is right; final text changes it	Cleanup strength and Exact replacements
Ordinary notes look fine; an AI input is rewritten more	“Write prompts in AI coding tools” and detected target
History has the right text; the field has none	Insertion, target focus and the copy fallback

A better microphone cannot repair text that was changed after recognition. Lowering cleanup strength cannot disable an explicit replacement rule either.

Pick the least editing that serves the task

Setting	Intended editing	Use it for
Verbatim	Punctuation with minimal wording changes	Wording you want to review closely
Standard	Fillers and explicit slips removed; spoken order retained	Messages, feedback and ordinary notes
Refined	Repetitions consolidated and supported relationships organized	A rambling brief or multi-part note

Lists → Never (prose only) controls layout. It does not turn Refined into Verbatim. Likewise, selecting Verbatim does not turn recognition into an exact record of the audio: recognition may already omit a filler or mishear a word. Local replacements, spoken-symbol handling and a selected Chinese script lock can also affect final text.

Cleanup labels differ across apps. For example, Wispr’s official Auto Cleanup guide separates raw, light cleanup and rewording. When comparing tools, compare what the setting permits and whether you can recover the original, rather than assuming “light” or “verbatim” means the same thing everywhere.

A worked example from a text-only backend check

On October 17, 2026, we sent this synthetic transcript to Ducduc’s production cleanup endpoint in Standard, with locale zh-CN. The request used no supplied vocabulary, protected terms or recent context. There was one attempt; the response below is unedited.

Exact input:

嗯这个 release 先别 deploy，我刚才说周五，不对改成下周一，先保留 3 个测试，不要删掉 unchanged 这个字段。

Actual output:

这个 release 先别 deploy，改成下周一。先保留 3 个测试，不要删掉 unchanged 这个字段。

The output removes the filler and explicitly retracted Friday, keeps next Monday and the 3 tests, and preserves both the “do not deploy” restriction and the instruction not to delete unchanged. It does not add an approval step or declare the release ready. The date phrase still needs to make sense to its recipient; cleanup cannot supply missing context.

The archived input, settings and response make this one check inspectable. This was text cleanup, not microphone dictation. It does not measure speech recognition, app detection, latency, repeatability or a general success rate.

A quick comparison should check:

• Entities and quantities: who receives it, what is being sent, dates, units and amounts.
• Negatives and conditions: “do not,” “only if,” “after review,” “leave unchanged.”
• Certainty: “might,” “probably” and unanswered questions.
• Scope: one button, one paragraph or the whole page.
• Language: embedded English remains English unless translation was selected.

A word such as “like” can be a hesitation or part of “I like this option.” Blindly deleting every occurrence is not a safe definition of filler removal.

Why a coding prompt can change more than a note

With Write prompts in AI coding tools enabled, recognized AI apps and supported AI website fields can receive instruction cleanup. A short request should become a short written instruction; a long brief can have its repeated requirements merged. This mode may reorganize more than Standard.

Illustrative input: “The settings button is small and I keep missing it. Make it easier to click, but don’t change the whole page.”

Illustrative result: “Make the settings button easier to click because its small size causes misclicks. Do not change the whole page.”

No dimensions, implementation method or extra test requirement should be invented. Ducduc prepares text; the receiving agent interprets and acts on it. Review before submitting. The Codex guide covers that workflow.

Target detection depends on information macOS exposes through Accessibility, including supported browser address information. It can fail to identify a field. It does not read the webpage or typed field content to decide the platform. If you want to compare cleanup strengths consistently, disable prompt conversion for that test.

What the checks and fallback protect

The current implementation checks some changed numeric values, lost negatives, missing protected terms and assistant-like answers. A rejected result falls back to the recognized wording, with applicable local output policies. These are narrow checks, not a semantic proof. Retractions, spoken numbers and intentional translation need different treatment, so a wrong result can still pass.

When cloud cleanup is unavailable, Ducduc can use a configured local backup in Hybrid mode; otherwise it preserves recognized text and labels the result. Built-in quota exhaustion also preserves text. Local Ollama and BYOK have separate setup requirements. A successful fallback does not establish that the original recognition was correct.

Keep a small review test

Test the worked example plus “Maybe move it to Friday,” “Do not clear the cache,” and one sentence with a personal name or amount. Compare raw and final text at each strength using the same model. Then inspect Dictionary → Exact replacements and learned corrections if the same unexpected substitution remains.

Built-in and cloud BYOK cleanup send text to the selected service. Same-Mac cleanup requires a loopback endpoint, a downloaded local Ollama model and disabled Ollama cloud features. Ducduc Offline mode disables its cloud cleanup route, not Ollama’s network access. Follow local setup; the data-flow guide explains the other boundaries.

Try Ducduc with Standard and a short note first. For a meaning-changing result, report a redacted raw/final pair, app and build, strength, list preference, prompt-conversion setting and relevant rule. That gives the developer a reproducible problem without requiring your whole private recording.