AI Text Polishing for Dictation: Prompts, Examples and Review
What is AI Text Polishing?
This guide covers dictation polishing in the OpenTypeless desktop app. AI text polishing asks a language model to edit an existing transcript. It may adjust punctuation, sentence structure, or formatting, depending on the prompt and model. It can also change a name, number, limitation, or intention incorrectly. Treat the result as an editable draft and compare it with the source before using it.
How the Pipeline Works
OpenTypeless first obtains a transcript through the selected STT route. When AI Polish is enabled and the LLM route has usable credentials or a Cloud session, the app can send that text to the selected LLM. Ordinary dictation can output raw text when polishing is disabled; tasks that require generation or selected-text editing need an available LLM. These paths are conditional, not a fixed step after every recording. Complete response time depends on the route and draft; this page provides no measured latency guarantee.
Compare LLM routes using the same draft
Groq
For Groq, measure the complete polishing response on short and long drafts; record the model and settings.
OpenAI / Claude
For OpenAI or Claude, compare tone, factual preservation, and terminology on the same draft and prompt.
DeepSeek
For DeepSeek, enter the current rate and estimate cost from your own monthly text volume before choosing it.
Ollama
A configured Custom Whisper-compatible endpoint handles the local or self-hosted STT route. Ollama handles only the separate local LLM polishing route. Privacy still depends on the endpoint, model dependencies, updates, logging, and actual network traffic, so verify the intended configuration before processing sensitive content.
Configure and verify the two processing routes
- Configure STT and verify a short recording first. Then configure an LLM route and enable AI Polish when you want it.
- Review the transcript and processed draft separately. Check which endpoint receives audio and which receives text; a local STT route does not make a hosted LLM local.
- Focus a disposable document and verify insertion and undo. A successful provider connection test does not verify the full microphone-to-document workflow.
A prompt that preserves the scope of the draft
Edit this transcript for punctuation and readable paragraphs.
Preserve names, numbers, negation, uncertainty, conditions, and technical identifiers.
Do not add facts, commitments, deadlines, or test results.
Keep ambiguous wording instead of silently choosing an interpretation.
Return an editable draft.These instructions express the desired behavior; they do not guarantee the model follows every constraint. Dictionary entries supply vocabulary in the polishing prompt. They do not train the speech model or guarantee exact names. Retest recurring terms with complete sentences.
Illustrative example: preserve uncertainty
Raw draft: i think we might ship on friday if the tests pass but do not promise the date yet
Desired edited draft: I think we might ship on Friday if the tests pass, but do not promise the date yet.
Incorrect edit to reject: We will ship on Friday.This example was written to explain review criteria. It is not a user testimonial or measured model output. The incorrect edit loses “might,” the test condition, and the instruction not to promise a date.
Illustrative example: keep names and numbers
Raw draft: please ask Maya to review ticket AB-104 and do not replace version 1.1.61 with the latest version
Desired edited draft: Please ask Maya to review ticket AB-104, and do not replace version 1.1.61 with the latest version.Evaluate the finished draft
- Use a short message, a technical paragraph, and a sentence containing a condition, a negation, and a number. Repeat the same samples for each candidate route.
- Record total waiting and correction time, missing limitations, changed identifiers, added facts, and insertion failures. Compare against typing or keeping the raw transcript.
- Choose by the results of your own workflow. Provider brand, a star rating, or a fluent sentence does not establish accuracy, speed, or lower cost.