In a recipe workflow, transcription turns an available signal—usually speech or visible text—into text. Extraction identifies recipe facts and organizes them into fields such as ingredients, quantities, and steps. Reconstruction supplies relationships or missing details that were not explicit. These labels are useful workflow distinctions, not a universal product taxonomy.

Transcription: what was said or written?

Examples:

  • speech “one quarter teaspoon” becomes text;
  • a caption is copied;
  • recognized text is read from an image.

Transcription can still be wrong. Fractions, accents, background music, overlapping speech, and ingredient names are common challenges. A correct transcript may also be incomplete because the speaker never states the full recipe.

Extraction: which parts form the recipe?

Extraction asks:

  • Which phrases are ingredients?
  • Which numbers are quantities, times, temperatures, or package sizes?
  • Which actions form steps?
  • What is the order?
  • Which ingredient does “the rest” refer to?

A source can be transcribed accurately but extracted incorrectly. “Bake in an 8-inch pan for 30 minutes” contains three different field types.

Reconstruction: what was not explicit?

Reconstruction or inference may turn visible or customary behavior into a completed-looking instruction. Examples include estimating an unspoken oil quantity or inserting a preheat step because the oven is later used.

That may be helpful as a suggestion, but it should not be represented as copied source fact.

One video can require all three

Suppose a creator says “add the rest of the milk,” shows a measuring jug, and lists “500 ml milk, divided” in the caption.

  • Transcription captures the spoken sentence.
  • Extraction identifies milk, total quantity, and a divided use.
  • Linking “the rest” to a prior step requires relationship reasoning.
  • Guessing the first portion when neither channel states it would be reconstruction.

Research on generating recipes from unsegmented cooking videos similarly treats event identification and instruction generation as a compound problem.

How to review a generated recipe

Mark each consequential detail as:

  • explicit in source;
  • structured from explicit source;
  • inferred or estimated;
  • missing;
  • your tested edit.

Mise extracts reviewable ingredients and steps from supported sources. Users should compare the result with the original and avoid treating an unsupported detail as source truth.

The important product question is not which marketing term appears on the feature. It is whether you can see, correct, and trace the result.

Try the workflow

Keep the source. Review the recipe. Cook from your version.

Download Mise on the App Store

Sources and further reading

  1. Recipe generation from unsegmented cooking videos
  2. NIST — automatic speech recognition
  3. BBC Good Food — how to write a recipe

Read how we research, draft, review, and correct guides in our editorial policy.