extended-wer-normalizer

jiwer-compatible text normalizer for Word Error Rate (WER) evaluation in voice AI.

Extends jiwer's built-in transforms with normalizations that matter for real-world ASR evaluation: phone numbers, emails, URLs, currency, percentages, ordinals, filler words, and stuttering.

Installation

pip install extended-wer-normalizer

Quick start

from extended_wer_normalizer import normalize_for_wer

normalize_for_wer("Call 0176 or email info@example.com, it costs $5.99")
# → "call 0 1 7 6 or email info at example dot com it costs five dollars ninety nine cents"

normalize_for_wer("Um, 1st place goes to Dr. Smith with 50% accuracy")
# → "first place goes to doctor smith with fifty percent accuracy"

jiwer integration

Every normalization is a jiwer.AbstractTransform subclass — compose them freely:

import jiwer
from extended_wer_normalizer.transforms import NormalizeEmails, ExpandDigitRuns

pipeline = jiwer.Compose([
    NormalizeEmails(),
    ExpandDigitRuns(),
    jiwer.ToLowerCase(),
    jiwer.RemovePunctuation(),
    jiwer.ReduceToListOfListOfWords(),
])

wer = jiwer.wer("info at example dot com", "info@example.com", hypothesis_transform=pipeline)

Use the pre-built pipeline directly with jiwer.wer:

import jiwer
from extended_wer_normalizer import english_wer_pipeline

wer = jiwer.wer(reference, hypothesis, reference_transform=english_wer_pipeline, hypothesis_transform=english_wer_pipeline)

Available transforms

Every transform that consumes language-specific data accepts a language="en" keyword (default English): NormalizeEmails(language="fr"), ExpandAbbreviations(language="de"), etc.

Pipeline design

The English pipeline applies transforms left-to-right in a single pass:

1. Pattern-specific (before punctuation is stripped): email, URL, symbol, abbreviation, currency, percentage, ordinal
2. Core: contractions (I'm → i am), lowercase, intra-word hyphen split (fixed-term → fixed term), punctuation removal
3. Digit normalization: expand digit runs (0176 → 0 1 7 6), convert digit words (zero → 0)
4. Cleanup: filler words, repetition collapse

Supported languages

Full pipelines (with language-specific abbreviations, fillers, lexicons, and number/ordinal/percentage word forms via num2words) ship for English, German, and French. Pass any other language value for the minimal fallback (lowercase + punctuation + whitespace).

from extended_wer_normalizer import normalize_for_wer

# German: titles, fillers, ordinals, currency
normalize_for_wer("Hr. Müller, am 1. Januar, ähm, ungefähr 50% Rabatt", language="de")
# → "herr müller am erste januar ungefähr fünfzig prozent rabatt"

# French: elision contractions, ordinals, comma-decimal currency
normalize_for_wer("M. Dupont, le 1er janvier, c'est €5,99", language="fr")
# → "monsieur dupont le premier janvier c est cinq euros quatre vingt dix neuf centimes"

# Spanish, Italian, … fall through to the minimal pipeline
normalize_for_wer("¡Hola, mundo!", language="es")
# → "hola mundo"

# Tune the repetition-collapse threshold (default 3 → collapse runs of 4+)
normalize_for_wer("no no no", max_repeats=1)  # collapse any 2+ run
# → "no"

Per-language pipelines are also exposed for direct use with jiwer.wer:

from extended_wer_normalizer import (
    english_wer_pipeline,
    german_wer_pipeline,
    french_wer_pipeline,
)

To inspect or extend the language data:

from extended_wer_normalizer.languages import get_language_data, supported_languages

supported_languages()              # ["de", "en", "fr"]
get_language_data("de").abbreviations["hr."]  # "herr"

Quirks worth knowing

• Hyphenated compounds: intra-word hyphens split into spaces (fixed-term → fixed term), so hyphenated and spaced spellings of the same words score identically. Before 0.6.0 the hyphen was deleted (fixedterm), which penalized STTs that never emit hyphens — WER numbers are not comparable across that boundary.
• Comma vs. period decimals: French uses , (€5,99, 3,5%); the currency and percentage transforms accept either separator regardless of language.
• German ordinals: matched as 1- to 3-digit numbers followed by . and a word (e.g. "1. Januar" but not "Es war 1990." or "1.5 Liter"). 4+ digits and decimals are skipped to avoid false positives on years.
• French ordinals: matched as 1er, 1ère, 2e, 2es, 2ème, 2èmes, 2nde, 2nds, 2nd. num2words returns masculine forms (premier, deuxième); feminine variants like première or seconde are not produced.
• Contractions: jiwer.ExpandCommonEnglishContractions runs only for English. French has a custom ExpandFrenchElisions that splits j', l', d', n', s', m', t', c', qu', jusqu', lorsqu', puisqu', quoiqu' from the following word. German has no contraction step.
• German pluralization: most currency units stay invariant (fünf Euro, not fünf Euros).