wiki / narration-tts-text-normalization / edit

Suggest an edit.

Improvements are reviewed before going live, like any wiki.

Title

One-line summary

The prompt

Act as the text normalization pass of a text-to-speech pipeline. I will give you raw text that may contain numerals, symbols, markup, URLs, and formatting. Return text a TTS engine can synthesize cleanly, plus a report of every decision that could change meaning.

Text: [PASTE RAW TEXT]
Language and locale for spoken forms: [e.g. ENGLISH (US)]
Maximum characters per output chunk: [ENGINE LIMIT, OR "NO LIMIT"]
Domain context for resolving ambiguity: [e.g. FINANCIAL NEWS, RECIPE, LEGAL DISCLAIMER]

Normalization rules:
1. Numbers: expand to spoken words per the locale — years ("twenty twenty-four"), currency ("four dollars and fifty cents"), decimals ("three point one four"), phone numbers digit by digit, ordinals, ranges ("five to ten"), and units ("kilometers per hour", never "km slash h").
2. Abbreviations: expand ("Dr." to "Doctor", "e.g." to "for example") unless the spoken form is the abbreviation itself. Acronyms: render letter by letter ("A P I") or as a word ("NASA") according to common usage; when unsure, choose one and log it.
3. URLs and emails: spoken form ("example dot com slash pricing") only if the listener needs the address itself; otherwise substitute a natural reference ("the link in the description") and flag the substitution.
4. Strip markdown, HTML, emoji, and decorative symbols. Convert tables and bullet lists into spoken sentences.
5. Split sentences longer than 25 words at clause boundaries. Preserve paragraph breaks — they carry pause information.
6. Chunking: if a limit is set, number the chunks and break only at paragraph or sentence boundaries, never mid-sentence.

Return, in order: the normalized chunks; a pronunciation dictionary of remaining risky tokens (names, jargon) with phonetic respellings; and an ambiguity log listing each token whose expansion required a judgment call ("2019" as year versus quantity, "St." as saint versus street), the reading you chose, and the alternative, so I can correct it before synthesis.

Write it so someone can paste it straight into Claude, ChatGPT or Gemini and fill the [BRACKETS].