8. Quirks, dead code & corrections

Things a faithful port has to reproduce, things a clean reimplementation should drop, and where earlier notes were wrong.

Bugs that shipped

Glyph scale → volume is a dead store

if (scale > 0x20) {
    scale = scale - 0x20;
    f31 = scale;
    f31 = f31 * 0.15f;
    f30 = sou_voice_se[0]._08 + f31;   // computed…
    f30 = 4.f;                          // …and immediately discarded
    sou_voice_se[0]._08 = f30;
} else {
    scale = 0x1f - scale;
    f31 = scale * 0.02f;
    f30 = sou_voice_se[0]._08 - f31;    // computed…
    f30 = 0.4f;                         // …and immediately discarded
    sou_voice_se[0]._08 = f30;
}

game64.c_inc:2994

The intent was clearly a smooth scale-to-volume ramp. What ships is a step function: any glyph larger than normal slams volume to 4.0 (a 6× jump over the 0.65 baseline), anything smaller to 0.4. Normal-size text passes scale == 32 and skips the block entirely, so this is only audible on deliberately scaled dialogue.

This is in the matching decomp, so retail does it. A faithful port needs it; a clean library does not.

The mode-2 dedup check is tautological

case 2: {
    if (sou_num_org == sou_num_org && sou_num_org == sou_last_num2) i = 1;
}

game64.c_inc:2711

The first comparison is always true. The intent was presumably a three-in-a-row check mirroring mode 0; the effect is a two-in-a-row check. Deferred syllables therefore trigger the random substitution more often than intended.

SHIIN2BOIN is dead, and would read out of bounds

const u8 SHIIN2BOIN[] = { 0, 1, 2, 3, 4, 5, 1, 2, … };   // 86 entries
…
sou_now_boin = SHIIN2BOIN[a];    // a can be up to 0x77 = 119

game64.c_inc:236, written at :2863

sou_now_boin is written here and initialised in Sou_InitAudio, and never read anywhere in the tree. Same for sou_nobasu_count ("elongation count"), which is only ever set to 0. Both are vestigial mouth-shape / vowel tracking, presumably live in the N64 Dōbutsu no Mori code this was ported from. Because the table is 86 entries and a can reach 119, the write reads past the array — harmless only because the result is discarded.

The table also has an obvious data oddity: entry 45 is the value 45, where every neighbour is 0–5.

Design choices that surprise

n is a vowel. Sou_BoinShiinCheck groups it with a e i o u. This changes when silent e and magic e fire, so done, gone and none get the silent-e drop.
h is silent. Sou_TanboinHenkan('h') returns 0x00, and code 0 has no instrument (guarded by if (a != 0 && a <= 0x77)). Standalone h makes no sound; it only speaks through ch sh th ha he hi ho oh.
A space can never open a window. mMsg_sound_voice_entry refuses voice_no == 0x81 in the "first character" branch, so word-initial pairing always starts on a letter.
Animalese is never panned or attenuated. pan stays at 0x3F and reverbVolume at 0 for the life of the program. A villager at the far edge of the screen is exactly as loud and as centred as one in your face — the only positional SE in the game that ignores position.
The two subtracks share one parameter set. sou_voice_se[0] and [1] are always written together, and Sou_VoiceMake pushes both every frame — so a new syllable retroactively changes the volume and pitch of the one still ringing.
Voice mode is not the config setting. Animalese only ever plays for messages that pass mMsg_check_sound_animal; and the submenu forces sAdo_SetVoiceMode(0) regardless of config, so letter-writing always speaks even for players who set the game to silent.
Sequence-load races drop characters. After a spec change, Na_VoiceSe returns early until Nap_ReadSubPort(VOICE_GROUP, 0, 1) == sou_now_voice_seq. The first syllables of the first line with a new speaker can be silently lost.

Corrections to earlier notes

The following supersede statements in ../extraction.md as originally written.

Was	Actually
"Bank 157 has only 83 instruments … that voice collapses every digraph sample above `0x53` onto one sound."	Nothing above `0x53` is a digraph. The phonology engine's maximum output is `0x52`, plus `0x77`. Bank 157 covers the phonology range exactly; the missing block is the editor's letter names, unreachable at spec 3. The only real loss is `you`.
"43 of the 119 codes are unreachable through normal text … only ever hit by the anti-repetition random substitution."	36 of them (`0x53`–`0x76`) are reachable — through the mode-1 editor path — and the random substitution can not reach them, since it only produces `0x00`–`0x3F`.
"`Sou_SpecialRoutine10` replays it on the next audio frame."	It runs inside `Na_GameFrame`, which `graph.c` calls once per game frame. Same ~17 ms at 60 fps, but the mechanism is the game loop, and it halves under `SetGameFrame(2)`.
"♪ … calls `Na_MelodyVoice(npc_id)` — the villager's tune."	It plays a fixed 16-note pattern `{13, 14 × 15}`, not a tune. The per-NPC part is which instrument: `npc_voice_id` selects a chunk of sequence 248 to copy in. It is the speaker's timbre playing a constant figure.
Specs 5 and 6 listed only in the spec→sequence table, unattributed.	They are the player's own voice in text-entry menus — 5 male, 6 female — set from `m_submenu.c:364`, and are special-cased throughout `Na_VoiceSe`.
Voice mode not covered.	Three modes gate the whole pipeline, and the gate depends on the message, not just the config. See gating.
Mode `f == 1` not covered.	A third calling convention that bypasses phonology entirely. See the editor path.

Everything else in extraction.md §1 was re-verified against the current tree and holds: the voice table, the 55 digraph entries, both vowel tables, the per-spec timbre values, and the spec→sequence→bank mapping.