Terminology

Package terminology adapts the canonical term IR from terminology-spec into the two text-transform primitives every TTS/subtitle consumer of OmniVoice needs. It is the single shared implementation of that substitution engine — consumers (videoascode and others) should use it instead of reimplementing a regex-based term matcher.

The two primitives

Primitive	When it runs	What it fixes
Pronouncer	Before TTS synthesis	The spoken form of a term, per BCP-47 locale — so the audio says "AAuth" as "ay auth"
CaseCorrector	After STT (subtitle text)	The written/displayed form of a term, independent of locale — so subtitles read "PlexusOne", not "plexus one"

The two are deliberately opposite: Pronouncer changes what a term sounds like without touching the displayed spelling; CaseCorrector changes what a term looks like in transcribed text.

Term sources

import "github.com/plexusone/omnivoice-core/terminology"

// Embedded generic-industry terms (~237: AI/ML, companies, dev tools,
// languages, frameworks, cloud/infra) — no filesystem access required.
terms, _ := terminology.Builtin()

// Hand-authored canonical terms with real pronunciation/translation data
// (e.g. protocol/standard names like "AAuth", "OAuth").
curated, _ := terminology.Curated()

// Builtin merged with Curated (Curated wins on ID collision).
all, _ := terminology.All()

// Or load a project's own canonical terms/*.json directory.
custom, _ := terminology.LoadDir("terms")

Usage

// Pronunciation: rewrite spoken forms before sending text to a TTS provider.
p := terminology.NewPronouncer(all)
spoken := p.Apply("Sign in with AAuth", "en-US") // -> "Sign in with ay auth"

// Case correction: fix displayed forms in STT-derived subtitle text.
c := terminology.NewCaseCorrector(all)
fixed := c.Apply("sign in with aauth")           // -> "sign in with AAuth"

Term is re-exported from terminology-spec, so consumers don't need a second direct import just to name the type.

Matching semantics

• Case-insensitive with word boundaries.
• Multi-word terms are matched before shorter overlapping ones, so a longer term isn't shadowed by a shorter one nested inside it.
• Terms with no entry for the requested locale are left unchanged.

See also

• Local TTS/STT Providers — where pronunciation is applied on the synthesis side.
• videoascode's Pronunciation guide — the config-driven consumer of Pronouncer.