g2pflow

Composable multilingual grapheme-to-phoneme pipelines for Python 3.10+.

g2pflow preserves alternative readings, complete pronunciation paths and the pronunciation groups within each path. Converters control word boundaries and can be combined in priority order for mixed-language text.

Install

pip install g2pflow
pip install "g2pflow[ja]"    # MeCab and UniDic
pip install "g2pflow[lstm]"  # ONNX inference for LSTM converters
pip install "g2pflow[all]"   # Both optional backends

These are the intended PyPI installation commands. Before publication, install this checkout with pip install -e . (or pip install -e ".[all,dev]").

The base package depends only on Pydantic. NumPy and ONNX Runtime are loaded for LSTM OOV inference; filelock is loaded only when UniDic needs downloading. The lstm and ja extras install these respective optional dependencies.

Quick start

This example needs no downloaded dictionaries or models:

from g2pflow import G2PPipelineConfig, build_pipeline_from_config

config = G2PPipelineConfig.model_validate({
    "preprocessors": [{"id": "lowercase"}],
    "converters": [{
        "id": "characters",
        "language": "en",
        "kwargs": {"mapping": {"h": ["HH"], "i": ["AY"]}},
    }],
})
pipeline = build_pipeline_from_config(config)
word = pipeline.convert("Hi", languages=["en"])[0]
assert word.text == "hi"
assert word.readings[0].paths[0][0].phonemes == ["HH", "AY"]

G2PPipeline also accepts preprocessor and converter instances directly. Configuration is validated by independent Pydantic models; there is no YAML loader, training configuration system or command-line interface.

PFML markup

PFML (Pronunciation Flow Markup Language) adds language scopes, fixed word boundaries and complete pronunciation candidate trees to ordinary text:

from g2pflow import G2PPipeline, to_pfml

# Direct phonemes need no dictionaries, models or converters.
words = G2PPipeline().convert_pfml(
    '<scope language="ja"><word script="ka" phonemes="k a">か</word></scope>'
)
assert words[0].readings[0].paths[0][0].phonemes == ["k", "a"]
assert G2PPipeline().convert_pfml(to_pfml(words)) == words

<scope language="ja">東京へ行く</scope> routes a whole text range to Japanese G2P; <word language="zh">重庆</word> fixes one output word. Automatic text requires a pipeline configured for the corresponding language. Direct results must provide all final phonemes; they preserve multiple readings, paths and group boundaries without invoking G2P or filtering candidates. Missing outer containers can be omitted, down to <phoneme>ong</phoneme>.

There is no required root tag. convert() still handles plain text, while convert_pfml() handles XML fragments. Python outputs remain G2PWord objects; to_pfml() serializes successful outputs losslessly, including all pronunciation candidates and language markers. See the PFML 1.0 specification and API.

Output

• G2PWord(text, language, readings) is a converter-defined semantic word.
• G2PReading(paths) groups the complete realizations of one reading.
• G2PPath is an ordered list[G2PGroup].
• G2PGroup(script, phonemes) associates a pronunciation script with phones.

For example, a reading of Japanese 猫 may contain a path with groups ("ne", ["n", "e"]) and ("ko", ["k", "o"]). Different paths can have different group boundaries. The framework filters silent groups and paths, omits words left without pronunciation, and fills missing display labels. Missing readings or path candidates are conversion failures. The package does not select candidates against audio or encode phones into model vocabulary IDs.

Components and resources

Preprocessors: filter-punctuation, strip-whitespace, lowercase, and remove-accents.

Mandarin and Cantonese character/phrase reading data are bundled. The caller chooses the phoneme inventory through external pronunciation dictionaries. English ONNX models are external. The Japanese backend manages UniDic's first-use download and can also use a preinstalled directory. See resources and converter details.

External plugins

from g2pflow import register_plugin_paths

register_plugin_paths("./my_plugins", "../shared_plugins")

Each directory may contain both preprocessors and converters, registered with @preprocessor and @converter. Registration scans immediately. Call discover_plugins() to find newly added modules; existing modules load once per process. Paths are supplied only through Python, not through configuration. See plugin discovery and multiprocessing.

Development

pip install -e ".[dev]"
pytest --cov=g2pflow --cov-report=term-missing
python -m build
python -m twine check dist/*

Default tests run offline with small fixtures and controlled backend doubles. Real-resource tests and clean-install distribution tests are explicit opt-ins; see development and validation.

Licensing

Project code uses the MIT license. Redistributed code and data retain their respective licenses; the distribution's license expression also includes Apache-2.0, CC-BY-SA-3.0 and CC-BY-SA-4.0. See third-party notices.