When down-sampling an audio time series to a lower sample rate, one must first remove frequency components above the Nyquist frequency of the new time series.  This is accomplished by low-pass filtering the data before resampling, a process sometimes referred to as "anti-aliasing".  Failure to anti-alias the time series results in high-frequency components being "aliased" down to lower frequencies, corrupting the audio.

The audioresample element does not anti-alias the time series when down-sampling.  In a pipeline, one can work-around this bug by preceding the audioresample element with a low-pass filter element like audiowsinclimit, but audioresample, like any element, should Just Work (tm).