So the implementations for PLP feature extraction I have seen in python and Matlab, are as follows:
suppose the FFT features of an audio signal are of the size 257 (512 point FFT). if we have one frame of the features: size=(257,1) and we implement a bark filterbank consisting of 20 filters: (size=20*257). then the output of this implementation is of size (20*257)*(257*1)=(20*1) which means that now we have 20 features instead of 257 for each frame. but when I checked this paper by Hynek Hermansky and a few other papers, It is stated that the filterbank is convoluted with the power spectrum. but what we did in the code wasn't convolution, was it? also aren't filters convoluted in the time domain or else multiplied by the signal in the frequency domain? I am pretty sure there is something about the mathematics of it that I am missing.