Removing Noise with AI - RM Noise

Hi Randy,

You wrote: "I’m pretty sure our method of determining the bandwidth of the audio is flawed. "

please check my gist, there is python code to generate band limited noise wav file, and link to soundspec python which generates spectrum image showing boundaries

73

Howdy.

Here are 3 examples that were "passed" as 2.8KHz:


Often it isn't exactly clear which is at 2.8KHz vs 3.0KHz. I *believe* this is because different radios have different DSP roll-off strategies between the passband and stopbands.


To illustrate the different DSP roll-off shapes, here are the 2'nd and 3'rd recordings:


Notice the radio's DSP roll-off is very "sharp" on the 3'rd recording.


Even with my own lowpass filtering at the server, the neural network performs differently with all of these different DSP shapes, and only through a variety of recordings in the training set does it perform well across them all.