meFFT
>>> from scipy.signal.windows import gaussian
...
>>> T_x, N = 1 / 20, 1000  # 20 Hz sampling rate for 50 s signal
>>> t_x = np.arange(N) * T_x  # time indexes for signal
>>> f_i = 1 * np.arctan((t_x - t_x[N // 2]) / 2) + 5  # varying frequency
>>> x = np.sin(2*np.pi*np.cumsum(f_i)*T_x) # the signal

The utilized Gaussian window is 50 samples or 2.5 s long. The parameter
``mfft=200`` in `ShortTimeFFT` causes the spectrum to be oversampled
by a factor of 4:

>>> g_std = 8  # standard deviation for Gaussian window in samples
>>> w = gaussian(50, std=g_std, sym=True)  # symmetric Gaussian window
>>> SFT = ShortTimeFFT(w, hop=10, fs=1/T_x, mfft=200, scale_to='magnitude')
>>> Sx = SFT.stft(x)  # perform the STFT

In the plot, the time extent of the signal `x` is marked by vertical dashed
lines. Note that the SFT produces values outside the time range of `x`. The
shaded areas on the left and the right indicate border effects caused
by  the window slices in that area not fully being inside time range of
`x`:

>>> fig1, ax1 = plt.subplots(figsize=(6., 4.))  # enlarge plot a bit
>>> t_lo, t_hi = SFT.extent(N)[:2]  # time range of plot
>>> ax1.set_title(rf"STFT ({SFT.m_num*SFT.T:g}$\,s$ Gaussian window, " +
...               rf"$\sigma_t={g_std*SFT.T}\,$s)")
>>> ax1.set(xlabel=f"Time $t$ in seconds ({SFT.p_num(N)} slices, " +
...                rf"$\Delta t = {SFT.delta_t:g}\,$s)",
...         ylabel=f"Freq. $f$ in Hz ({SFT.f_pts} bins, " +
...                rf"$\Delta f = {SFT.delta_f:g}\,$Hz)",
...         xlim=(t_lo, t_hi))
...
>>> im1 = ax1.imshow(abs(Sx), origin='lower', aspect='auto',
...                  extent=SFT.extent(N), cmap='viridis')
>>> ax1.plot(t_x, f_i, 'r--', alpha=.5, label='$f_i(t)$')
>>> fig1.colorbar(im1, label="Magnitude $|S_x(t, f)|$")
...
>>> # Shade areas where window slices stick out to the side:
>>> for t0_, t1_ in [(t_lo, SFT.lower_border_end[0] * SFT.T),
...                  (SFT.upper_border_begin(N)[0] * SFT.T, t_hi)]:
...     ax1.axvspan(t0_, t1_, color='w', linewidth=0, alpha=.2)
>>> for t_ in [0, N * SFT.T]:  # mark signal borders with vertical line:
...     ax1.axvline(t_, color='y', linestyle='--', alpha=0.5)
>>> ax1.legend()
>>> fig1.tight_layout()
>>> plt.show()

Reconstructing the signal with the `~ShortTimeFFT.istft` is
straightforward, but note that the length of `x1` should be specified,
since the SFT length increases in `hop` steps:

>>> SFT.invertible  # check if invertible
True
>>> x1 = SFT.istft(Sx, k1=N)
>>> np.allclose(x, x1)
True

It is possible to calculate the SFT of signal parts:

>>> N2 = SFT.nearest_k_p(N // 2)
>>> Sx0 = SFT.stft(x[:N2])
>>> Sx1 = SFT.stft(x[N2:])

When assembling sequential STFT parts together, the overlap needs to be
considered:

>>> p0_ub = SFT.upper_border_begin(N2)[1] - SFT.p_min
>>> p1_le = SFT.lower_border_end[1] - SFT.p_min
>>> Sx01 = np.hstack((Sx0[:, :p0_ub],
...                   Sx0[:, p0_ub:] + Sx1[:, :p1_le],
...                   Sx1[:, p1_le:]))
>>> np.allclose(Sx01, Sx)  # Compare with SFT of complete signal
True

It is also possible to calculate the `itsft` for signal parts:

>>> y_p = SFT.istft(Sx, N//3, N//2)
>>> np.allclose(y_p, x[N//3:N//2])
True

Ú