NeMo Inverse Text Normalization: From Development To Production

Boris Ginsburg; Evelina Bakhturina; Kyle Gorman; Yang Zhang

NeMo Inverse Text Normalization: From Development To Production

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2104.05055 v2 pith:7IVOD5HS submitted 2021-04-11 cs.CL cs.SDeess.AS

NeMo Inverse Text Normalization: From Development To Production

Yang Zhang , Evelina Bakhturina , Kyle Gorman , Boris Ginsburg This is my paper

classification cs.CL cs.SDeess.AS

keywords textnormalizationlibrarydevelopmentinversenemooutputproduction

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

0 comments

Share X Bluesky LinkedIn Reddit HN

read the original abstract

Inverse text normalization (ITN) converts spoken-domain automatic speech recognition (ASR) output into written-domain text to improve the readability of the ASR output. 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. We introduce an open-source Python WFST-based library for ITN which enables a seamless path from development to production. We describe the specification of ITN grammar rules for English, but the library can be adapted for other languages. It can also be used for written-to-spoken text normalization. We evaluate the NeMo ITN library using a modified version of the Google Text normalization dataset.

discussion (0)

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Length-Aware Rotary Position Embedding for Text-Speech Alignment
eess.AS 2025-09 conditional novelty 6.0

Length-aware RoPE (LARoPE), which normalizes positional indices by sequence length, induces a diagonal attention bias that improves text-speech alignment and achieves state-of-the-art WER on a zero-shot TTS benchmark.
Earnings25: A Comprehensive 500-Hour Speech Benchmark for Finance
cs.CL 2026-07 accept novelty 4.0

Earnings25 releases ~500 hours of 2025 earnings-call audio with aligned transcripts, speaker/industry metadata, and reproducible Whisper and Parakeet-TDT baselines.