Hector Martin

That is *exactly* what happened. Well done! It was some kind of confusion in the interface (a power cycle fixed it) that caused it to drop every second sample and replace it with a 0. The aliasing wasn't entirely obvious due to the audio I was using.

I noticed when I turned on spectrogram view for 'B' and saw the mirror image spectrum... which wasn't *completely* implausible given 'B' was chiptune stuff recorded digitally, but bizarre... then I zoomed in to the PCM and saw every other sample was 0.

It wasn't strictly speaking up/downsampling since everything involved was nominally 48kHz (so I was getting aliased 24kHz audio), but same effect.

was I right that the recording had made the same error

Yeah, that was why 'B' was special like that. 'A' was just some random stuff from another source. And the (digital) level meter of the interface was before the point in the pipeline where the sample dropping happened, so it always showed activity.

This actually got me thinking of the DSP implications. I know how sampling and aliasing work in the frequency axis, but clearly phase matters too. Shifting the sampling phase shifts the phase at which the upper spectrum half folds over? I need to think about this.

Like, obviously when the downsampling aligns with the nonzero samples you get constructive aliasing interference, and when it aligns with the zeros you get destructive interference, but why?

Paging @xiphmont. Any pointers as to how to explain this time-domain phenomenon in the frequency domain? How does the phase of aliased downsampling affect the phase of the spectral components that get folded together?