Port of krom (Peter Lemon)'s SpeechSynth (PeterLemon/SNES, SPC700/SpeechSynth). Speech on the SNES = sequencing a phoneme bank on one DSP voice: per phoneme, the APU-side program selects the sample (SRCN), shapes it (ADSR: plosives one-shot, vowels/nasals looped and sustained), inflects pitch for prosody, and waits (4 ms timer grain). krom's 32-phoneme bank spells "PETER LEMON"; our original 5-phoneme bank (formant synthesis, gen_phonemes.py) says "OPEN SNES" on loop.

ROM mode: LoROM (project default).

Register fidelity vs the original

Element	krom	this port
Master	MVOL 127/127, echo fully off (FLG $20)	same
Voice 0	VOL 127/127, GAIN 127	same
Vowel ADSR	%11110111/%11111100	same values
Plosive ADSR	%11111111/%11100000	same values
Pacing	32 ms consonants / 256 ms vowels	same (4 ms tick × 8/64)
Retrigger	bare KON per phoneme, no KOFF	same
Phrase end	plain wait, no KOFF (ADSR decays)	same

The debugging archaeology (all owner-listening-test driven)

Flat-binary layout collision: the sequencer outgrew $0200-$02FF; directory moved to $0800.
Timer 2× too fast: T0DIV=255 gave ~16 ms effective ticks in practice; krom's 1 ms grain replicated at 4 ms — vowels finally last their 256 ms (vowel-segment count now matches the reference 9 vs 8).
The "toc": an inter-phrase KOFF pulse we added truncated the last phoneme mid-waveform; krom lets the ADSR decay naturally. Removed.
The perceptual verdict: with all fixes, the owner A/B'd our machinery against krom's demo in BOTH luna and Mesen2 — the "noises/defects" heard are present in the original too: that IS what robotic phoneme spelling sounds like. Port machinery verified by luna spc-dump end-to-end: uploads byte-intact, directories entry-equivalent, DSP registers equal, phoneme schedule correct.

Debug tooling that cracked it

luna spc-dump exports the full APU state (64 KB ARAM + all 128 DSP registers) as a playable .spc — diffing two dumps localized every one of the bugs above. See docs/tutorials/audio.md.

Modules Used

console, apu