Skip to main content
Inverse text normalization converts spoken-form text produced by speech recognition into the written forms that people and software expect.

​
ITN and text normalization

Text normalization prepares written text for speech synthesis. It can turn $20 into twenty dollars. Inverse text normalization handles the opposite direction after speech recognition.

​
Why rules are useful

Deterministic rules are good at validating structure. They can ensure that a date is possible, preserve leading zeroes in an identifier, or render a phone number consistently.

​
Why context matters

The same words can represent different values:
A rule can generate both interpretations. Sentence context is needed to choose the intended one. In the retained room-code example, Premove ITN outputs the room code is 105. The benchmark also records ambiguity errors, so this is not a guarantee that every contrast is resolved correctly.

​
Contextual ITN

Premove ITN combines deterministic candidate generation with contextual ranking. Rust generates valid written forms, DeBERTa scores them in the full sentence, and an exact decoder selects compatible edits.

​
Why it matters after ASR

ASR transcribes what was said. ITN decides how a spoken value should be written. For a person, zero eight two zero six three is readable. For a database key, preserving the leading zero and producing 082063 can be necessary for a successful lookup. The correct formatting also depends on what the sentence means. A time, an amount, and a room number can share spoken words. ITN does not repair transcription errors or infer a value that the speaker did not say. It also does not replace application-side validation. A voice agent should still check values before a consequential API action. Read why voice agents need ITN, how Premove ITN works, or install the package.