Possible off-by-one bug in number of output samples

[kylebgorman]

Thanks for scikits.samplerate: it looks like a great package. I am, however, surprised at the number of samples I get from calling resample on a mono signal: it is always exactly one less than my expectations. I wanted to confirm this was the correct behavior and not a bug before I go on using the package "in production". Here’s a worked example that shows what I mean:

from __future__ import division
from math import trunc
from numpy.random import uniform
from scikits.samplerate import resample

L = 10000
SR_in = 44100
SR_out = 16000

signal_in = uniform(size=L)
signal_out = resample(L, SR_in / SR_out, 'sinc_best')
expected_length = trunc(L * SR_in / SR_out)
print expected_length - 1 = len(signal_out) # prints “True"

I took a look at scikits.samplerate-0.3.3/scikits/samplerate/_samplerate.pyx. If I understand correctly, on line 173

osz = (r * nframes - 1)

sets the output length. That explains why the resampled audio is the length it is (original length times ratio, minus 1, truncated towards zero because of the cast to a long integer), but why is 1 being subtracted here?

It’s quite possible there’s a legitimate signal processing reason for this. But if so, what is the actual samplerate of signal_out?