buchwandler

Prepared phonemization

KokoroG2P is the phonological and model-adaptation layer. Its core APIs consume prepared, speakable text plus an explicit language. They preserve that text as the coordinate space for tokens, annotations, overrides, and results.

from kokorog2p import phonemize_prepared

result = phonemize_prepared("Version two point oh.", language="en-us")

The core does not verbalize numbers, abbreviations, units, currencies, dates, times, URLs, or version labels. Applications that need those semantics must prepare the text first, for example with the optional Spokenform package, and then call phonemize_prepared().

Linguistic annotations

Applications can inject analysis results without making KokoroG2P import the analysis package. Objects need start, end, and may provide text, pos, tag, lemma, and language.

from kokorog2p import TokenAnnotation, phonemize_prepared

annotations = [
    TokenAnnotation(0, 6, "record", pos="NOUN", tag="NN", lemma="record")
]
result = phonemize_prepared(
    "record this record", language="en-us", annotations=annotations
)

Offsets are half-open source offsets. They must be ordered, non-overlapping, within the supplied text, and an annotation’s optional text must match its source slice. When annotations are supplied, the pipeline uses them for POS metadata and does not run the configured spaCy model for that call.

Explicit language routing

Language switching is explicit. Use OverrideSpan or annotation language metadata for a foreign span. The core still requires an explicit document/default language and does not automatically select that document language.

Optional automatic pronunciation-language routing can inspect selected lexical resources for individual words or exact sub-token fragments:

result = phonemize_prepared(
    text,
    language="de",
    language_routing={"mode": "auto", "languages": ["de", "en"]},
    overlap="split",
)

The candidate list is a hard allowlist. Lexicon collisions and ambiguity stay in the default language. Generic pronunciation fallback is not language evidence, and explicit ph, phonemes, lang, and language overrides take precedence.

Lexphon 0.1.3 supplies clean generic IPA and structured pronunciation language markers. For German-default DE/EN routing, a marker such as en is lexical evidence that can authorize a bounded pair-specific loanword route when the English candidate is unique and compatible. It does not globally override German ownership: unmarked whole-token collisions remain German/default. KokoroG2P uses the structured marker metadata and never parses raw (en) or (de) source syntax.

Migration from pre-v0.9

Remove semantic preparation flags such as input_mode, migrated_semantics, and enable_context_detection. Remove abbreviation registry calls and number/abbreviation expansion options. Prepare the source text in the owning application or cross-package pipeline, then pass the prepared result to phonemize_prepared() with its explicit language.