I’m getting really tired of the artificial sound and formants imprinted on granulated sounds due to the windowing functions used for the grains. I’ve been listening to these sounds a lot lately as I keep working on gRainy Days the Jitter based granulation system that started of as a port of CloudGenerator. At the moment I’m desperate at trying various window functions to see if it’s possible to minimize artefacts. How the different windows behave probably depends a lot on the sound material I’m working on as well as what densities and grain sizes I’m using.
Some of the windows have additional problems when used for granulation as they don’t reach zero at the end points. This might introduce digital clicks. For synthetic granulation the clicks might only appear at the end of the window (as a sinus function starting of at zero is used). For granulation of audio buffers the clicks might appear at the beginning and the end of the window. CloudGenerator uses a gaussian curve scanning the function from -3 to 3.000 Hence the gain at the end points of the window is down by ca. 20 dB only compared to the middle point. When working on granulation of audio buffers CloudGenerator has very pronounced artefacts with a strong dependency on the density of the grains. I suspect that this is not only due to spurious frequencies and formants introduced by the window in itself but also clicks at start and end points introducing frequencies of their own (for instance if using 300 synchonized grains per second I risk a possible artificial frequency of 300 Hz with a rich set of harmonics).
I haven’t seen this issue raised in discussions on what windows to use for granulation but it’s obvious that if e.g. Gaussian or Hamming is used they should be modified at the ends to ensure that they reach zero properly.
I’ve kept adding window functions to the tl_window java class. So far these windows are implemented: Hamming Hanning Bartlett (Triangle) Blackman BlackmanHarris Gaussian Kaiser Square Sinc CosineAlpha CosineTaper and Trapezoid.
