Render the reduced de-/re-/pre- prefix vowel as ə, not ᵻ

[christopherthompson81]

Reported as "determine sounds tensed". Not a dictionary defect — our IPA is already reduced on every path — but a divergence between two spellings of the same reduced vowel, which matters because Kokoro was trained on one of them. In misaki's us_gold the prefix vowel is ⟨ə⟩ 1,057 times against ⟨ᵻ⟩ 9, so our ⟨ᵻ⟩ is effectively out of distribution for this position.

Established by synthesis, not by the distribution argument

An A/B through Kokoro itself confirmed the model renders the two differently (rms 0.05–0.08; determine comes out 50 ms longer with ⟨ᵻ⟩ against identical leading silence). A listener then preferred ⟨ə⟩ on a connected-reading sample and called the difference "pretty subtle, overall, in this reading" — which is the size of claim this change gets to make, and it's written into the code and log so nothing downstream inflates it.

The sample was built as the candidate fix rather than a blanket swap, so the listen answered the actual decision:

changed  determine  dᵻtˈɜɹmən  ->  dətˈɜɹmən
changed  describe   dᵻskɹˈIb   ->  dəskɹˈIb
changed  reduce     ɹᵻdˈus     ->  ɹədˈus
KEPT     deduce     dᵻdˈus                     (ded- carve-out)

The rule

ReduceEnglishPrefixVowel, walking groups against the group → word map. Three conditions, each earned:

1. The source word begins de-/re-/pre- — the word, not the phonemes.
2. Except ded- — the deduce/deduct family plus two surnames is the only place gold attests a prefix ⟨ᵻ⟩, i.e. the words with the strongest evidence in the class. Verified upstream: 0 of the 1,057 target words begin ded-, and 0 of the keep-words don't.
3. The ⟨ᵻ⟩ must be the token's first vowel. represent is ɹˌɛpɹᵻzˈɛnt, whose ⟨ᵻ⟩ is a second syllable with nothing to do with the prefix.

It does nothing when the map is null or mismatched: a word-keyed rule that has lost the word must leave the stream alone rather than guess.

⚠ Why word-keyed is the whole point. Earlier drafts tried to express this phonologically and needed a dᵻd key plus a tie-bar guard, because dᵻd͡ʒ (degeneracy, deject) matches a bare d on half of /d͡ʒ/. All of that was manufactured by working at a layer without word identity. PhonemizeTrace is the full word → IPA trip and Phonemize already holds its map, so the carve-out is orthographic — ded- is exact, deg-/dej- are excluded by spelling, and the affricate hazard never arises.

⚠ dejected → dəʤˈɛktᵻd is the case to keep in mind: the prefix moves and the inflectional ⟨ᵻ⟩ — gold's own spelling, 1,287 entries — stays, in one word. Any rule phrased as "replace ⟨ᵻ⟩" fails exactly there.

Tests, and what they can't catch

Reverted the rule and ran them: 7 target-class cases go red, and the 10 boundary cases pass either way. That's correct and worth saying rather than glossing — boundary cases assert things are unchanged, so they can't detect the rule's absence. Their job is catching over-application, which is the failure that would actually damage a reading.

One expectation had to be weakened: dedans is rare enough to go through the neural OOV path, so its tail is dᵻdˈæns or dᵻdˈænz depending on whether that model resolved — the exact expectation was testing the model's presence, not this rule. Exact-IPA expectations for OOV words are unstable by construction, which applies to any future test in this file.

Suite: 95 + 271 + 23 + 362, 0 failures.

Not fixed by this

• The phonemizer still emits ⟨ᵻ⟩ in prefix position, so any other consumer of the library inherits the divergence untreated. An argument for eventually fixing it upstream too — recorded so "we fixed it" isn't later read as "it is fixed". vernacula-phonemizer#1445 holds the measurements either way.
• be- (before, become, begin) shows the same ⟨ᵻ⟩ and is out of scope: nobody has measured gold's spelling for that onset, so it's left alone rather than swept in for looking similar.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KfTCTjC7FTnNm3aS2zszKE

[christopherthompson81]

Review pass — four findings, all fixed in 7852e44

1. The rule disagreed with Render about what English is. KokoroFormat.Render's English arm is lang is "en" or "en-GB" or "en-US" or null; my gate omitted null. A call with a null language would render English phonemes and then skip the English rule on the same call. Now aligned to Render's own definition rather than to a list I wrote from memory.

2. map[group++] was unbounded. The count check above it should make an overrun impossible, but this runs on every synthesised paragraph and an index bug would take the utterance down rather than degrade it. Out of range now means what it already means everywhere else in this rule — cannot identify the word, leave the token alone.

3. KokoroVowels.ToCharArray() allocated on every prefix word. Static array.

4. The tests didn't check the output was still in vocabulary. ⚠ audio.cpp refuses a supplied stream carrying a symbol Kokoro has no id for — it doesn't drop it, it rejects the request — so a rule emitting one would take out synthesis on that backend while the ONNX path silently skipped it. Asserted per case now.

Confirmed rather than assumed: both engines get this

KokoroTts (ONNX) calls Phonemize(text, british), and Lang(british) returns "en"/"en-GB". Had the rule reached only one backend, the same document would read differently depending on the engine — the exact failure KokoroChunker exists to prevent.

⚠ And my control was invalid, for the second time this week

First attempt stashed only the uncommitted review fixes while the rule itself was already committed in 1b63a79 — so the "control" ran against the rule and reported 17 green, making the tests look vacuous when they weren't.

I wrote this exact trap up on #240 and then walked into it again. Redone against main:

7 failed, 10 passed

The #240 lesson was "check what the control actually reverted". The stronger version: stash is the wrong instrument whenever any part of the change is committed. Reverting a named file to main doesn't depend on remembering which half is in the index.

Verification

Suite: 95 + 271 + 23 + 362, 0 failures. 7 target-class tests red without the rule; the 10 boundary tests pass either way, as they should — they guard against over-application, not absence.