WeTextProcessing Runtime

Python runtime for WeTextProcessing (does not depend on Pynini).

WeTextProcessing is a text processing library that provides text normalization (TN) and inverse text normalization (ITN) capabilities for Chinese, English and Japanese text. It uses Finite State Transducers (FSTs) for efficient text processing.

Features

• Text Normalization (TN) for Chinese, English and Japanese
• Inverse Text Normalization (ITN) for Chinese, English and Japanese
• Traditional to Simplified Chinese conversion
• Full-width to Half-width character conversion
• Interjection removal
• Punctuation removal
• Out-of-vocabulary (OOV) word tagging
• Erhua removal (for Chinese)
• 0-to-9 conversion (for Chinese and Japanese ITN)

Installation

pip install wetext

Usage

Python API

Text Normalization (TN)

from wetext import Normalizer

# Chinese TN with erhua removal
normalizer = Normalizer(lang="zh", operator="tn", remove_erhua=True)
result = normalizer.normalize("你好 WeTextProcessing 1.0，全新版本儿，简直666")
print(result)  # 你好 WeTextProcessing 一点零，全新版本，简直六六六

# English TN
normalizer = Normalizer(lang="en", operator="tn")
result = normalizer.normalize("The price is $12.50, please pay now.")
print(result)  # The price is twelve point five dollars, please pay now.

Inverse Text Normalization (ITN)

from wetext import Normalizer

# Chinese ITN
normalizer = Normalizer(lang="zh", operator="itn", enable_0_to_9=False)
result = normalizer.normalize("你好 WeTextProcessing 一点零，全新版本儿，简直六六六，九和六")
print(result)  # 你好 WeTextProcessing 1.0，全新版本儿，简直666，九和六

# English ITN
normalizer = Normalizer(lang="en", operator="itn")
result = normalizer.normalize("twenty three dollars and fifty cents")
print(result)  # $23.50

Streaming TN and ITN

StreamNormalizer keeps unfinished normalization context internally. Pass only the new text on each call; feed() returns the complete current output. Use one instance per utterance and call flush() when the utterance ends.

from wetext import StreamNormalizer

stream = StreamNormalizer(lang="zh", operator="itn")
print(stream.feed("今天是二零"))  # 今天是二零
print(stream.feed("二六年"))  # 今天是2026年
print(stream.flush())  # 今天是2026年

stream = StreamNormalizer(lang="zh", operator="tn")
print(stream.feed("价格是12"))  # 价格是十二
print(stream.feed(".5元"))  # 价格是十二点五元
print(stream.flush())  # 价格是十二点五元

Streaming requires an explicit language and supports TN and ITN. Model bundles built before streaming prefix graphs were introduced remain compatible; they retain more internal context until flush(). For an unusually long unfinished expression, only its tail is normalized for the current display; flush() always returns the exact normalization of all remaining input.

N-best and exact mappings

# Distinct candidates in official WFST weight order
normalizer = Normalizer(lang="en", operator="tn")
print(normalizer.normalize("4x6", nbest=3))
# ["four by six", "four times six", "four x six"]

# Include tropical WFST costs (lower is better; costs are not probabilities)
candidates = normalizer.normalize_candidates("4x6", nbest=3)
print(candidates[0].text, candidates[0].cost)

# Exact token spans traced through the selected WFST path
normalizer = Normalizer(lang="zh", operator="itn")
result = normalizer.normalize_with_mapping("价格是十三点五元")
print(result.output_text)  # 价格是¥13.5
print(result.mappings[0].as_dict())

normalize_with_mapping() cannot be combined with fix_contractions=True, because contraction expansion is not represented by the WFST path.

Command Line Interface

# Basic usage
wetext "你好 WeTextProcessing 1.0，全新版本儿，简直666"

# With options
wetext --lang zh --operator tn --remove-erhua "你好 WeTextProcessing 1.0，全新版本儿，简直666"

# Convert traditional to simplified Chinese
wetext --traditional-to-simple "你好，這是測試。"

# Remove punctuations
wetext --remove-puncts "你好，這是測試。"

API Reference

Normalizer Class

Normalizer(
    lang: Literal["auto", "en", "zh", "ja"] = "auto",
    operator: Literal["tn", "itn"] = "tn",
    traditional_to_simple: bool = False,
    full_to_half: bool = False,
    remove_interjections: bool = False,
    remove_puncts: bool = False,
    tag_oov: bool = False,
    enable_0_to_9: bool = False,
    remove_erhua: bool = False,
)

Parameters

• lang: The language of the text. Can be "auto", "en", "zh" or "ja". Default is "auto".
• operator: The operator to use. Can be "tn" (text normalization) or "itn" (inverse text normalization). Default is "tn".
• traditional_to_simple: Whether to convert traditional Chinese to simplified Chinese. Default is False.
• full_to_half: Whether to convert full-width characters to half-width characters. Default is False.
• remove_interjections: Whether to remove interjections. Default is False.
• remove_puncts: Whether to remove punctuation. Default is False.
• tag_oov: Whether to tag out-of-vocabulary words. Default is False.
• enable_0_to_9: Whether to enable 0-to-9 conversion for ITN. Default is False.
• remove_erhua: Whether to remove erhua for TN. Default is False.

Methods

• normalize(text: str, lang: Optional[Literal["auto", "en", "zh", "ja"]] = None) -> str: Normalize the text.

CLI Options

• --lang, -l: Set the language. Choices are "auto", "en", "zh", "ja". Default is "auto".
• --operator, -o: Set the operator. Choices are "tn", "itn". Default is "tn".
• --traditional-to-simple: Convert traditional Chinese to simplified Chinese.
• --full-to-half: Convert full-width characters to half-width characters.
• --remove-interjections: Remove interjections.
• --remove-puncts: Remove punctuation.
• --tag-oov: Tag out-of-vocabulary words.
• --enable-0-to-9: Enable 0-to-9 conversion.
• --remove-erhua: Remove erhua.

License

Apache License 2.0