Yuval did the following:- Selected an mp3 file for the music...

Yuval did the following:
- Selected an mp3 file for the music demo using the software displayed in the video, which displayed the analog audio signal for the whole mp3 (took ~1 sec to appear).
- Selected a ~40sec section of the mp3 (due to memory limitations on the FPGA on the donut) and then normalised the signal (<1 sec).
- Powered up the donut and connected to it.
- Selected the settings for the algorithm to process the signal (including the 5Bit setting), then processed the signal and uploaded it over ethernet to the FPGA's memory (~1 min in video time but he edited that part so probably actually took ~2 mins).
- Played the audio (i.e. the processed audio signal on the FPGA was sent to the chip) and recorded the output from the speaker at 48khz.

"The sound output from the MEMS was then sampled by the microphone and displayed on the screen as a static display trace." - correct. The recorded signal displayed in the software was too small to see any details (he didn't zoom in). It didn't sound marvellous as you point out for good reason (echoing my thoughts it seems).

He then repeated this process for the specific tones/sweep he demo'd (except they're available directly within the software so no mp3 file to select), which processed a lot faster for obvious reasons.