Tempo algorithm is wrong.

[enn-nafnlaus]

Tempo appears to be calculated as "duration / (# of beats - 1)". That's not an accurate algorithm, as not all beats will be detected (or even present in the song), while half-beats will sometimes be detected.

A proper algorithm is:

tempo = sorted([beat_times[i+1] - beat_times[i] for i in range(len(beat_times) - 1)])[int((len(beat_times) - 1) / 2)]

or

tempo = statistics.median([beat_times[i+1] - beat_times[i] for i in range(len(beat_times) - 1)])

... depending on whether you want a hard median (no averaging in the case of there being an odd number of detected beats, aka an even number of gaps between them) or a soft median (with averaging between the two centre elements).