buchwandler

Language profiles

Profiles add sparse language-specific knowledge to the generic IPA feature distance. A profile may specify an intended language/variety, a version, descriptive inventory metadata, notation aliases, low-cost sequence equivalences, and provenance sources.

Profiles should not become:

  • pronunciation dictionaries

  • G2P engines

  • TTS-specific rewrite systems

  • exhaustive N×N distance matrices

Bundled profiles

The currently bundled profile set is:

de-DE

The aliases de, de-DE, and de_de resolve to canonical profile ID de-DE. Profile IDs and aliases are normalized case-insensitively.

from phonodist import available_profiles, get_profile

assert available_profiles() == ("de-DE",)
assert get_profile("de").id == "de-DE"
assert get_profile("de-DE").id == "de-DE"
assert get_profile("de_de").id == "de-DE"

Using a profile

from phonodist import available_profiles, get_profile, pronunciation_distance

print(available_profiles())
profile = get_profile("de_de")
result = pronunciation_distance("n̩", "ən", language=profile.id, explain=True)

LanguageProfile.inventory is descriptive metadata and an audit aid, not an exhaustive rejection list. PanPhon remains authoritative for whether a segment has a feature representation in metric v1. The inventory does not enumerate every length mark, narrow allophone, loanword segment, or external G2P realization. See the LanguageProfile API entry for profile fields.

Rule types

Alias

Use an alias when two strings are alternate notation for the same intended segment in the profile.

de-DE: t͜s -> t͡s

Aliases are reserved for representation-level canonicalization. Generic normalization already canonicalizes the supported tie-bar variant; untied affricates are modeled as sequence equivalences so their zero-cost alignment remains visible in explanations.

Sequence equivalence

Use a sequence equivalence when common language-specific realization or phonological behavior maps different segment sequences closely.

de-DE: ə n ↔ n̩

These rules have low non-zero cost so the explanation remains visible.

Inventory semantics

The bundled inventory is descriptive profile metadata and an audit aid. It is not an exhaustive rejection list in metric v1. PanPhon remains authoritative for whether an IPA segment has a feature representation, because the profile does not enumerate every length mark, narrow allophone, loanword segment, or external G2P realization.

Provenance

Profile authors should cite linguistic/research sources, but should not copy incompatibly licensed linguistic databases into Apache-2.0 profile files.