Improve dictation language and punctuation
Transcription produces spoken language written down. That is not the same as written language, and knowing the difference makes dictated prompts much more useful.
Language
Speak in whichever language you would write the prompt in. The transcript comes back in the language you spoke — dictation transcribes, it never translates.
The generation tools work in the site's language, taken from the site locale. Dictating in one language and generating in another works, but the prompt then arrives in a language different from the output, which produces less reliable results than a prompt written in the site's language.
On a multilingual site, dictate in the language of the site you are working on.
Switching languages mid-sentence
Mixing languages in one recording is where transcription struggles most. A Spanish sentence containing three English product names is a normal way to speak and a hard thing to transcribe.
Expect those inserted terms to come back approximated. Fix them in the prompt field rather than re-recording.
Punctuation
Punctuation is inferred from your pauses and intonation. It is usually reasonable at sentence level and unreliable at anything finer.
What you can rely on:
- sentence breaks, mostly, when you pause naturally;
- basic capitalisation.
What you cannot:
- commas in long sentences;
- colons, semicolons, dashes and brackets;
- quotation marks;
- paragraph breaks.
Speaking punctuation aloud — saying "comma", "full stop" — is not a supported command. Those words are transcribed literally, which is worse than the missing punctuation.
The practical approach: speak naturally, accept approximate punctuation, and fix it in the field. For a prompt, punctuation matters less than clarity anyway.
What transcription gets wrong most
Proper nouns. Product names, brand names, company names and people's names are the least predictable part of any transcript, because they are exactly the words a general transcriber has least reason to expect.
Industry vocabulary. Specialist terms, abbreviations and acronyms.
Numbers. Prices, quantities, dates, percentages, model numbers and phone numbers. Always check these, because a generator will repeat a wrong figure with complete confidence.
Homophones. Words that sound identical are resolved by context, and dictated prompts are often short on context.
Anything spelled out letter by letter. Reference codes and SKUs rarely survive.
For all of these the fix is the same: type them instead of dictating them. Dictate the explanation, type the identifiers.
Speaking for a good transcript
- Get close to the microphone. Distance costs more accuracy than an accent does.
- Reduce background noise. Music and nearby conversation degrade results sharply.
- Speak at a normal pace. Slowing down unnaturally does not help and often hurts.
- Pause between sentences. This is what gives the transcriber its punctuation cues.
- Plan before you start. A recording of you deciding what to say transcribes as you deciding what to say.
- Use a headset if you dictate often.
Clean up the narration
Dictated prompts carry speech habits that a written prompt would not:
- false starts and self-corrections — "no, sorry, I mean…";
- filler — "so", "basically", "you know";
- repetition, because speech repeats more than writing;
- length, because spoken sentences run longer.
All of it is transcribed faithfully, and all of it becomes part of the instruction the generator reads. Deleting it takes seconds and improves both the result and the cost, since the prompt is part of what an operation is priced on.
A workflow that plays to its strengths
Dictation is good at getting a lot of rough material down quickly, and poor at precision. Use it that way:
- Dictate the situation, the audience and what the piece should cover.
- Delete the false starts.
- Type the precise parts yourself — names, numbers, exclusions.
- Tighten it into instructions rather than narration.
- Estimate and generate.
The result is a prompt with the completeness of speech and the precision of writing, which is better than either alone.
Do not dictate sensitive information
Everything transcribed becomes prompt text, and prompt text leaves your infrastructure. Never dictate credentials, customer records or anything confidential — the rules are identical to typed prompts. See Keep sensitive information out of prompts and Audio privacy.