Lexphon

Lexphon is a generic, lexicon-driven phonemizer and CLI built on G2Lex. It consumes released pronunciation assets and returns normalized IPA without producing Kokoro phonemes.

g2lex-data producer -> G2Lex storage -> Lexphon runtime -> application adapter

The ownership boundary is deliberate:

• g2lex-data obtains, transforms, validates, licenses, and publishes immutable data releases.
• G2Lex stores and queries typed lexicon data.
• Lexphon explicitly installs, verifies, selects, and consumes released assets.
• KokoroG2P or another application converts generic IPA to its model-specific output.

Lexphon must never contain source acquisition or G2Lex build recipes for production dictionaries.

Package layout

There is intentionally no src/ directory. The import package lives at ./lexphon.

Install for development

python -m pip install -e ".[dev]"
pytest

Data management

Catalog access and downloads are explicit. The user-visible data states are distinct: a catalog declaration is not proof of remote publication, and remote publication is not local installation. DataStore.install() is the network-capable provisioning operation. Phonemizer, lookup, token phonemization, rendering, and local store inspection never fetch a catalog or download a lexicon.

lexphon data available lists artifacts declared by the selected catalog only. It does not download the referenced manifest or asset or verify that release files are reachable. lexphon data install retrieves and verifies those files. If a catalog points to a missing release resource, the install diagnostic reports the logical ID, manifest or asset, release tag, data version, URL, and HTTP or connectivity failure, and confirms that the failed lexicon was not installed.

lexphon data available de-DE
lexphon data install de-de:gold
lexphon data list
lexphon data info de-de:gold
lexphon data verify de-de:gold
lexphon data remove de-de:gold
lexphon phonemize --language de-DE "Die Leute kommen."

When upgrading from older built-in defaults, provision the new generic assets explicitly before using lexicons=None:

lexphon data install en-us:lexhint
lexphon data install en-gb:lexhint
lexphon data install fr:lexhint
lexphon data install th:lexhint-native

Applications that intentionally need an older *:gold application-specific asset should select it explicitly through their downstream integration.

Use --catalog PATH_OR_URL and --data-home PATH for a local release or alternate store. Installation downloads the manifest first, verifies manifest and asset hashes and sizes, checks catalog and manifest identity, opens the G2Lex asset, and atomically activates a complete version. A copied, pre-populated store eliminates Lexphon catalog and lexicon downloads during runtime. Optional providers have separate provisioning requirements. The production German assets are de-de:gold, de-de:crane, de-de:espeak, and de-de:olaph. English CMUdict is available as the explicit ARPABET alternative en-us:cmudict. LexHint logical IDs use :lexhint for the preferred/default asset and :lexhint-native for an explicit native-Wiktionary alternative. Native-only languages such as Thai use only the native suffix; English's en-us:lexhint and en-gb:lexhint use native English Wiktionary.

Catalog pronunciation encodings are not all generic Phonemizer alphabets. Lexphon can discover, install, and verify application-specific assets such as kokoro-v1, but selecting one as a generic Phonemizer layer raises UnsupportedAlphabetError. Kokoro vocabulary conversion remains in KokoroG2P.

Data release versions and the Lexphon Python package version are independent. Pin the data catalog or release during provisioning, and pin the Python dependency separately.

CLI

lexphon --help
lexphon languages
lexphon phonemize --language de-DE "Die Leute kommen."
lexphon phonemize --language de-DE --lexicon de-de:crane "Die Leute kommen."
lexphon phonemize --language de-DE --lexicon de-de:crane --tag DET "die"
lexphon phonemize --language en-US --lexicon en-us:cmudict --json "read"

JSON output has schema_version: 2 and contains the rendered IPA plus structured token fields: text, pronunciation, source category, provider, requested language, source encoding, logical lexicon ID, matched key, structured variants, selector tag, known status, punctuation status, and provenance metadata.

Inline language-control markers recognized in IPA source notation are consumed once at the Lexphon normalization boundary. They are removed from pronunciation and each variant pronunciation, while source_pronunciation and structured language_markers preserve the raw source and marker metadata for diagnostics and optional downstream routing. Optional generic provider use is explicit:

lexphon phonemize --language de-DE --fallback espeak "unbekannteswort"

Install optional providers separately when using them:

python -m pip install "lexphon[espeak]"
python -m pip install "lexphon[goruut]"

The eSpeak extra does not install the system executable; provide eSpeak or eSpeak-NG through the operating system. Goruut/Pygoruut may provision its own runtime, so a pre-populated Lexphon data store does not by itself guarantee fully offline provider startup.

Providers are disabled by default. Unknown tokens remain visible to downstream applications.

Python API

from lexphon import DataStore, Phonemizer

store = DataStore()
with Phonemizer(
    "de-DE",
    lexicons=["de-de:gold"],
    store=store,
    fallback=None,
) as g2p:
    result = g2p.phonemize_tokens("Die Leute")
    for token in result.tokens:
        print(token.text, token.pronunciation, token.source, token.lexicon_id)

phonemize_tokens() is the integration API. It preserves token-level provenance, selectors, structured variants, punctuation, and unknown words. IPA is normalized to Unicode NFC. ARPABET and CMU-style pronunciations are converted deterministically to IPA. Unsupported alphabets and invalid pronunciation tokens raise stable Lexphon exceptions.

PronunciationToken.pronunciation is derived from the first structured variant. source_pronunciation and language_markers expose its raw source notation and provenance. Downstream adapters should consume this metadata and must not search returned pronunciations for raw tags such as (en) or (de).

lookup_lexicon(token, tag=...) searches only lexicons and returns None on a miss, so it is safe for lexical evidence and language routing. lookup(token, tag=...) searches the lexicons first and then invokes the configured generic provider. Configure fallback=None, fallback="espeak", fallback="goruut", or a custom PronunciationProvider.

For example, a provider result such as (en)fˈIl(de) is returned as clean fˈIl while retaining source="provider", provider="espeak", requested_language="de-de", the raw source_pronunciation, and markers en@0 and de@4.

Batch lookup and provider contracts

lookup_many() preserves input order and checks every configured lexicon before invoking a provider. A provider receives only the missing tokens, and a runtime-checkable BatchPronunciationProvider can handle those misses in one call:

results = engine.lookup_many(["one", "two", "three"])

When fallback="espeak", the eSpeak provider is batch-capable: lookup_many() invokes eSpeak once for the batch of lexicon misses rather than once per token.

Batch output must contain exactly one str or None result per submitted miss. Lexphon rejects strings, bytes, non-sequences, wrong cardinality, and malformed elements with ProviderOutputError. Provider execution failures raise ProviderExecutionError; None remains a genuine direct miss. All provider output is normalized at the engine boundary.

Lexicon configuration

lexicons=None uses the language profile's default lexicons. Construction raises LexiconNotInstalledError when a required default asset is not installed. lexicons=[] explicitly selects no lexicon layers and is the provider-only configuration. Use the empty list in provider-only applications and tests when profile defaults should not be opened.

KokoroG2P boundary

KokoroG2P should import Lexphon's Python API and configure Lexphon's generic providers when needed. It converts returned IPA using its model-specific vocabulary and applies its own stress, ratings, and diagnostics policy. Lexphon does not import KokoroG2P, perform Kokoro validation, or download dictionaries during phonemization.