Quoted from G-P-E:I don't see how you can definitively call what appears to be a spike for a clock to be a software error.
The first two Black Knight instruction descriptions I posted above explain how the 6821 CB2 signal is pulsed for the CVSD clock signal (taken from the 6821 datasheet). The pulse ultimately takes a period of one CPU E clock. A 3.58MHz crystal is hooked up to the 6802 CPU which internally divides it by 4 resulting in the E clock signal having a period of about 1.1us. That's the time length of the CB2 low going pulse to clock the CVSD in the Black Knight waveform shown above.
This is an unforeseen implementation issue for the 55536/55564 but like you said, poor Harris datasheets probably didn't help.
barakandl now has working software fixes based on manually changing the PIA CB2 state like Jungle Lord. Interestingly, Gorgar which was the first Williams game with speech is what Jungle Lord CVSD code was based on. So they did it correctly to start with.
Black Knight and Alien Poker were likely done by a different programmer to the other speech games as these two also implemented the ability to change voice pitch, both of which have the issue with 55536/55564.
The challenges are that Black Knight has no free ROM space, and the instruction cycle count needed to be maintained in order to keep the voice pitch the same since the software loop for spitting out the data stream to the CVSD is CPU execution timed. But it was managed.
Another minor software error they made was at the very start of any spoken word, the clock signal is pulsed *before* the first valid speech bit is applied to the CVSD data input.