Preemphasis generates a glitch in the start of the output signal

The current implementation of the preemphasis generates a glitch in the output signal.

Steps to reproduce:

a = np.array([9.12e-05, 9.15e-05, 9.13e-05, 6.10e-05, 3.05e-05, 3.07e-05, 6.10e-05])
b = librosa.effects.preemphasis(a)
print(b)

Out: 
array([-9.699088e-01,  3.036000e-06,  2.545000e-06, -2.756100e-05,
       -2.867000e-05,  1.115000e-06,  3.122100e-05])

Here b[0] is clearly off.

What can be done is to asign b[1] to b[0]. This hack would deal with the glitch while keeping the lengths of b and a the same.

[gerazov]

Awesome 😎 keep up the good work guys!

[lostanlen]

Yes, considering that values of coef are typically very close to 1, i would still support linear extrapolation. Replicating y[0] (“constant”) or y[1] (“reflect”) would break the continuity of the first derivative, and thus create potential artifacts in the emphasized output.