We have some previous recordings of the various games. I don't even remember where in previous threads on the forum I'd found this, but I may as well attach it here now...
To repeat what's been said in the past: if we change the clock of the µPD775x part, we should be able to get enough more data such that we can generate a perfect 9-bit back-translation of the DAC values. But this recording I found to be just a little too noisy for me to be confident I could calculate that exact values used in the DAC.
The datasheet for the part implies each recording inside the IC can choose between dividers of 80, 104 (edit1: determined from spectrogram of above recording), or 128 to generate the sample clock (640kHz→8, 6.2, or 5kHz). Although normal 44kHz sound card sampling "should" be enough, I'd really be happier with a larger amount of oversampling ... and ideally using a sampler that
doesn't have a highpass filter on its input.
(edit2:
MAME's implementation of the related µPD7759 thinks that waveforms in the ROM can pick sample rates that are any integer divisor of 160kHz, and change this rate at any arbitrary time, including in the middle of sample playback. They also say that divisors smaller than 9 might not work, and oddly enough say that the slowest divisor is 32 while they mention enough bits to support up the 64.)
The parts other than the µPD77P56 lack the required pins.