One Dictionary, One Owner

"Some words the model just reads wrong. A pronunciation dictionary is where you correct them — once, in one place, owned by the engine."

The Words That Come Out Wrong

Every voice model has terms it fumbles: acronyms it reads as words, names it stresses oddly, tool names and brand names with no obvious pronunciation. A pronunciation dictionary is the override layer — a mapping from a term to how it should actually be said. It's how CWK gets spelled out instead of read as a syllable, and how a name that kept coming out wrong finally comes out right.

Bellows Owns It; cwkPippa Proxies

Here's the ownership call, and it's the same shape you've seen all quest. Bellows is the single ground truth for pronunciation — the dictionary lives in the engine, edited through the engine's own editor. cwkPippa's admin surface does not keep a second copy that could drift; it proxies to Bellows. One dictionary, one owner, many surfaces reaching it. A term's correct pronunciation has exactly one home, so it can never be right in one app and wrong in another.

A Dictionary Change Is a Cache Change

Because a pronunciation affects the sound, the dictionary's version is folded into the cache identity. Fix how a word is said and every cached entry that used the old pronunciation is correctly missed — the next request re-synthesizes with the new sound rather than serving the stale one forever. The dictionary editor even offers a quick fix right where you Speak, so correcting a mispronunciation and hearing the corrected result is one short loop.

Code

One home for pronunciation; other surfaces proxy to it·python

# The dictionary overrides how specific terms are read. Bellows OWNS it.
DICTIONARY = {
    "CWK":    "C W K",         # spell it out, don't read it as a word
    "Pippa":  "PIP-uh",        # a name the model kept getting wrong
    "mkcert": "M K cert",
}

# cwkPippa's admin surface keeps NO copy of its own -- it proxies here:
#   cwkPippa Admin  ->  (proxy)  ->  Bellows dictionary   [single ground truth]
#
# The dictionary VERSION rides in the cache identity. Fix a pronunciation and
# every affected cache key changes -- the old audio is correctly missed.

Exercise

List three terms a text-to-speech system (or a coworker reading your code aloud) reliably gets wrong — an acronym, a name, a tool. Write a dictionary entry for each: the term and how it should be said. Then answer the ownership question: if two apps both need these corrections, where does the dictionary live so they can never disagree?

Hint

The pronunciation is the easy part. The architecture question is where the correction lives so a fix applies everywhere at once. One owner plus proxies beats two copies plus a sync you'll forget to run.