Preprint

NeMo Inverse Text Normalization: From Development To Production

Yang Zhang et al., arXiv, 2021 · DOI 10.48550/arXiv.2104.05055

Read on October 17, 2026

What it is: An open-source Python library from NVIDIA for inverse text normalization, the step that converts spoken-domain speech recognition output into written text, built on hand-written weighted finite-state transducer grammars for English. The paper describes how the same library moves from development to production and evaluates it on a modified Google text normalization dataset.
How much weight it carries: An arXiv preprint, not yet peer-reviewed in this form, from the team that builds the NeMo library. It is a primary description of one widely used open-source ITN system. It describes the library's own design and evaluation, not an independent comparison with other ITN systems.

Facts taken from this source

Inverse text normalization converts the spoken-domain output of a speech recognizer into written-domain text, to make the recognizer's output easier to read.
“Inverse text normalization (ITN) converts spoken-domain automatic speech recognition (ASR) output into written-domain text to improve the readability of the ASR output.” (Abstract)
Many state-of-the-art inverse text normalization systems use hand-written weighted finite-state transducer grammars, because the task has extremely low tolerance to unrecoverable errors.
“Many state-of-the-art ITN systems use hand-written weighted finite-state transducer(WFST) grammars since this task has extremely low tolerance to unrecoverable errors.” (Abstract)