This test was written in such a way that having a wrong data type
for the mp3 samples wasn't detected. Instead of using
np.frombuffer(dtype=int16), just do arithmetic directly on the
samples. During testing time we don't care if it might be a little
slower or use a little more RAM than ulab, and we don't care
whether it's actually an RMS calculation.  Just that it's
consistent and shows the audio data is correct, including its
defined data type.