I 'm reading the following article from NVIDIA which involves some audio processing a the beginning : link just here (part 3.2)
I don't understand the following part :
In practice, we use 520ms worth of audio as input, i.e., 260ms of past and future samples with respect to the desired output pose. We chose this value to capture relevant effects like phoneme coarticulation without providing too much data that might lead to overfitting. The input audio window is divided into 64 audio frames with 2x overlap, so that each frame corresponds to 16ms (256 samples) and consecutive frames are located 8ms (128 samples) apart. For each audio frame, we remove the DC component and apply the standard Hann window to reduce temporal aliasing effects. Finally, we calculate K = 32 autocorrelation coefficients to yield a total of 64×32 scalars for the input audio window. Although much fewer autocorrelations, e.g. K = 12, would suffice to identify individual phonemes, we choose to retain more information about the original signal to allow the subsequent layers to also detect variations in pitch.
What does a 2x overlap mean ? And also what are these samples they are talking about ?