urdunorm

Urdu text normalization and WER/CER scoring for ASR evaluation, under named protocols you can report. Zero runtime dependencies. pip install urdunorm

The problem, in five lines

from urdunorm import wer

reference = ["کے لیے شکریہ"]  # what the transcriber wrote
hypothesis = ["کیلئے شکریہ"]  # what the model emitted, compound joined

wer(reference, hypothesis)  # 0.0
wer(reference, hypothesis, mode="lums-compat")  # 0.667

Same audio, same words, one spelling difference, two published protocols. That gap is why Urdu ASR results do not compare across papers, and why this package makes you name the protocol.

What it does to text, under eval

The last row matters most: phonemic characters are a contract, and the package refuses to import if a rule table would merge one.

How it differs from urduhack and LughaatNLP

Those are general Urdu NLP toolkits: tokenizers, POS tagging, NER. This is an evaluation protocol, normalization plus scoring, nothing else. Different jobs.

Differences are measured, not asserted. This table is generated by compare_normalizers.py --emit-table running each package at a pinned version:

Measured on urdunorm 0.1.0, urduhack 1.1.1, LughaatNLP 1.3.1.

Both also carry heavy dependencies: urduhack imports TensorFlow at package import; LughaatNLP requires torch, transformers, scikit-learn and scipy.

Modes

Use eval unless you know why not. display is for showing transcripts to people. lums-compat reproduces WER We Stand (COLING 2025), the only published Urdu ASR benchmark, so numbers can be compared against it. Its five-character punctuation class keeps the Urdu full stop: that is theirs, not a bug here. Report under two protocols; a single-protocol Urdu WER compares to nothing.

Scoring

from urdunorm import wer, cer, score, normalize, EVAL

normalize("کیلئے", mode="eval")
wer(references, hypotheses, mode="eval")
cer(references, hypotheses, mode="eval")

result = score(references, hypotheses, mode="eval")
str(result)  # 'WER 12.40% [eval]'
EVAL.with_(repair_spaces=False)  # an ablation is a field override

Both sides are normalized under the named mode and the result records which protocol produced it. Tokenization is a whitespace split. The corpus rate is total errors over total reference tokens, never the mean of per-utterance rates, which would weight a three-word clip like a thirty-word one.

More

Why this exists · Measured comparison · Where the rules come from · Review the Urdu rules (no Python needed) · Contributing

The space-repair lexicon and the number tables have not been signed off by a native speaker, so numbers_to_words ships off in every mode.

Citing

@software{urdunorm,
  title  = {urdunorm: Urdu text normalization for reproducible ASR evaluation},
  author = {Hassan Bhatti, Hunzalah},
  year   = {2026},
  doi    = {10.5281/zenodo.21864104},
  url    = {https://github.com/hunzed/urdunorm}
}

Apache-2.0.