John Muradeli
John Muradeli
Above demonstrates `librosa` failing to produce `Sx.shape[1] == (len(x) - 1) // hop_len + 1`, so it's about verifying that formula. Example: ```python win_length, hop_length, n_fft, len(x) == 5, 3,...
The issue is having different rationale for framing; under my scheme, the idea is to **center `window` at each sample of the original sequence `x`, starting at sample 0, strided...
I see. Indeed the framing logic is identical between `librosa` and `ssqueezepy`, and the only difference's in padding; I've designed it specifically to achieve the described effect. If interested, the...
Indeed a convolution-like padding support is viable. I'll draw consideration toward one scheme in particular: zero-padding. The problem with most padding schemes is imposing some assumption on boundary behaviors, which...
By "standard formulation" I mean standard mathematical formula, which is indeed this:  Pasting relevant excerpt from [here](https://dsp.stackexchange.com/a/72590/50076); `modulated=False` is `librosa`'s behavior. "Modulated" (modified) sums well both directly and via...
Key plots:  This concerns any application working with _parts_ of `Sx` rather than whole, including reconstruction, ridge extraction, and reassignment. Adjacent rows have similar values but opposite signs, and...
Magnitude exactly the same indeed, and any methods that operate on magnitude won't see a difference. Reconstruction (of a real input), however, deals with the real part, and reassignment with...
1. Implementing a `util.frame(.., modulated=True)` so values are seated correctly the first time, then `fftshift(fft_window)`, seems most efficient. As for in-place multiply, firstly `arange(n_fft//2) -> arange(n_fft//2+1)`, and the sign's right,...
Nicer example with pure tone and a time-localized window; real & imag components follow a bell and are consistent in sign along columns: code ```python import numpy as np import...
We'd only fftshift `fft_window`, which is a 1D array, not `y_frames`, which is `n_hops` times bigger in size. The window can also be spared of a copy, sure. Idea is,...