There was a nice article on the beatport tech blog. I didn't manage to find it when I searched yesterday. Basically what they suggested was applying a low pass filter to isolate the base drums, then count the most common intervals between peaks (basically some kind of transformation) and then reconstruct the BPM from there by applying a range from 80-160 or similar. Hope this gives you a rough idea.

I implemented and tested it and it worked quite well with electronic music.

I did find this too, it's the same library used by most DJ software: https://www.sonicapi.com/docs/api/analyze-tempo

Works pretty well but requires to upload a file, then process it to get the BPM. So you would need smaller audio samples as the soundcloud stream is too big and takes ages.

Total shot in the dark but I don't suppose it was reading a 44.1kHz file at 48kHz or the other way round?