scipy 1.16.1__cp314-cp314-musllinux_1_2_x86_64.whl

scipy/signal/tests/test_short_time_fft.py

@@ -0,0 +1,1098 @@
+"""Unit tests for module `_short_time_fft`.
+
+This file's structure loosely groups the tests into the following sequential
+categories:
+
+1. Test function `_calc_dual_canonical_window`.
+2. Test for invalid parameters and exceptions in `ShortTimeFFT` (until the
+    `test_from_window` function).
+3. Test algorithmic properties of STFT/ISTFT. Some tests were ported from
+   ``test_spectral.py``.
+
+Notes
+-----
+* Mypy 0.990 does interpret the line::
+
+        from scipy.stats import norm as normal_distribution
+
+  incorrectly (but the code works), hence a ``type: ignore`` was appended.
+"""
+import math
+from itertools import product
+from typing import cast, get_args, Literal
+
+import numpy as np
+import pytest
+from scipy._lib._array_api import xp_assert_close, xp_assert_equal
+from scipy.fft import fftshift
+from scipy.stats import norm as normal_distribution  # type: ignore
+from scipy.signal import check_COLA, get_window, welch, stft, istft, spectrogram
+
+from scipy.signal._short_time_fft import FFT_MODE_TYPE, \
+    _calc_dual_canonical_window, closest_STFT_dual_window, ShortTimeFFT, PAD_TYPE
+from scipy.signal.windows import blackman, gaussian, hamming, nuttall, triang
+
+
+def test__calc_dual_canonical_window_roundtrip():
+    """Test dual window calculation with a round trip to verify duality.
+
+    Note that this works only for canonical window pairs (having minimal
+    energy) like a Gaussian.
+
+    The window is the same as in the example of `from ShortTimeFFT.from_dual`.
+    """
+    win = gaussian(51, std=10, sym=True)
+    d_win = _calc_dual_canonical_window(win, 10)
+    win2 = _calc_dual_canonical_window(d_win, 10)
+    xp_assert_close(win2, win)
+
+
+def test__calc_dual_canonical_window_exceptions():
+    """Raise all exceptions in `_calc_dual_canonical_window`."""
+    # Verify that calculation can fail:
+    with pytest.raises(ValueError, match="hop=5 is larger than window len.*"):
+        _calc_dual_canonical_window(np.ones(4), 5)
+    with pytest.raises(ValueError, match=".* Transform not invertible!"):
+        _calc_dual_canonical_window(np.array([.1, .2, .3, 0]), 4)
+
+    # Verify that parameter `win` may not be integers:
+    with pytest.raises(ValueError, match="Parameter 'win' cannot be of int.*"):
+        _calc_dual_canonical_window(np.ones(4, dtype=int), 1)
+
+def test_closest_STFT_dual_window_exceptions():
+    """Raise all exceptions in `closest_STFT_dual_window`."""
+    with pytest.raises(ValueError, match="Parameters `win` and `desired_dual` are.*"):
+        closest_STFT_dual_window(np.ones(4), 2, np.ones(5))
+    with pytest.raises(ValueError, match="Parameters `win` and `desired_dual` are.*"):
+        closest_STFT_dual_window(np.ones((4, 1)), 2, np.ones(4))
+    with pytest.raises(ValueError, match="Parameter win must have finite entries!"):
+        closest_STFT_dual_window(np.ones(4)*np.nan, 2, np.ones(4))
+    with pytest.raises(ValueError, match="Parameter desired_dual must have finite.*"):
+        closest_STFT_dual_window(np.ones(4), 2, np.ones(4)*np.nan)
+    with pytest.raises(ValueError, match="Parameter hop=20 is not an integer.*"):
+        closest_STFT_dual_window(np.ones(4), 20, np.ones(4))
+    with pytest.raises(ValueError, match="Parameter hop=2.0 is not an integer.*"):
+        # noinspection PyTypeChecker
+        closest_STFT_dual_window(np.ones(4), 2., np.ones(4))
+
+    with pytest.raises(ValueError, match="Unable to calculate scaled closest dual.*"):
+        closest_STFT_dual_window(np.ones(4), 2, np.zeros(4), scaled=True)
+
+
+@pytest.mark.parametrize("scaled", (True, False))
+@pytest.mark.parametrize('sym_win', (False, True))
+@pytest.mark.parametrize('hop', (8, 9))
+@pytest.mark.parametrize('m', (16, 17))
+@pytest.mark.parametrize('win_name', ('hann', 'hamming'))
+def test_closest_STFT_dual_window_roundtrip(win_name, m, hop, sym_win, scaled):
+    """Do roundtrip, i.e., compare dual of dual windows.
+
+    The  default for parameter `desired_dual` is also verified.
+    """
+    win = get_window(win_name, m, not sym_win)
+    d1, s1 = closest_STFT_dual_window(win, hop, np.ones_like(win), scaled=scaled)
+    d2, s2 = closest_STFT_dual_window(win, hop, scaled=scaled)  # equals d1, s1
+    d3, s3 = closest_STFT_dual_window(d1, hop, win * s1, scaled=True)  # roundtrip
+
+    # Validate default for parameter `desired_dual` works (hard coded => assert ok):
+    xp_assert_equal(d2, d1, err_msg="Default for parameter `desired_dual` is not ok!")
+    assert s2 == s1, "Default for parameter `desired_dual` is not ok!"
+
+    res = np.finfo(win.dtype).resolution * 5
+    xp_assert_close(s1*s3, 1., atol=res, err_msg="Invalid Scale factors")
+    xp_assert_close(d3, win, atol=res, err_msg="Roundtrip failed!")
+
+    if scaled:  # check that scaling factor is correct:
+        d3, _ = closest_STFT_dual_window(win, hop, np.ones(m) * s1, scaled=False)
+        xp_assert_close(d3, d1, atol=res, err_msg="Roundtrip failed!")
+
+
+@pytest.mark.parametrize('scaled', (False, True))
+def test_closest_STFT_dual_window_STFT_roundtrip(scaled):
+    """STFT Roundtrip correctness of closest dual window. """
+    np.random.seed(5613830)
+    n, m_num, hop = 15, 7, 3
+    w, desires_dual = hamming(m_num), triang(m_num)
+    d, _ = closest_STFT_dual_window(w, hop, desires_dual, scaled=scaled)
+
+    SFT = ShortTimeFFT(w, hop=hop, dual_win=d, fs=1, scale_to=None, phase_shift=None)
+    x = 10 * np.random.randn(n)
+    Sx = SFT.stft(x)
+    y = SFT.istft(Sx, 0, n)
+
+    atol = np.finfo(w.dtype).resolution * 10
+    xp_assert_close(y, x, atol=atol, err_msg="Invalid closest window")
+
+
+@pytest.mark.parametrize('scaled', (False, True))
+def test_closest_STFT_dual_window_STFT_roundtrip_complex(scaled):
+    """STFT Roundtrip correctness of closest dual window with complex values. """
+    np.random.seed(34905436)
+    n, m_num, hop = 15, 7, 3
+    win = 1j*hamming(m_num) + nuttall(m_num)
+    desires_dual = 1j*triang(m_num) + blackman(m_num)
+    dual, s = closest_STFT_dual_window(win, hop, desires_dual, scaled=scaled)
+
+    SFT = ShortTimeFFT(win, hop=hop, dual_win=dual, fs=1, fft_mode='twosided',
+                       scale_to=None, phase_shift=None)
+    x = 10 * np.random.randn(n) + 10j * np.random.randn(n)
+    Sx = SFT.stft(x)
+    y = SFT.istft(Sx, 0, n)
+
+    atol = np.finfo(win.dtype).resolution * 10
+    xp_assert_close(y, x, atol=atol, err_msg=f"Invalid complex closest window ({s=})")
+
+
+@pytest.mark.parametrize("win_name, nperseg, noverlap, scale_fac", ([
+                         ('boxcar',      16,        8,       1/2),
+                         ('boxcar',      18,       12,       1/3),
+                         ('boxcar',      16,       12,       1/4),
+                         ('bartlett',    16,        8,        1.),
+                         ('bartlett',    16,       12,       1/2),
+                         ('bartlett',    30,       25,       1/3),
+                         ('hann',        16,        8,        1.),
+                         ('hann',        18,       12,       2/3),
+                         ('hann',        16,       12,       1/2),
+                         ('blackman',     9,        6,     50/63),
+                         ('boxcar',       8,        7,       1/8)]))  # hop = 1
+def test_closest_STFT_dual_window_cola(win_name, nperseg, noverlap, scale_fac):
+    """Test if `closest_STFT_dual_window` generalizes `check_COLA`.
+
+    The parameters were taken from the `check_COLA` documentation.
+    Note that `check_COLA` only guarantees the existence of a dual window with constant
+    values but not that those values are unity (which is clear, when investigating the
+    'boxcar' examples). The values for `scale_fac` were determined empirically.
+    """
+    desired_dual = get_window(win_name, nperseg, fftbins=True)
+    assert check_COLA(desired_dual, nperseg, noverlap), "COLA cond. violated!"
+
+    win = np.ones(nperseg)  # check scaled window:
+    d_s, s = closest_STFT_dual_window(win, nperseg-noverlap, desired_dual, scaled=True)
+
+    res = np.finfo(desired_dual.dtype).resolution
+    rel_tol_d = max(abs(d_s))*res*3
+    xp_assert_close(s, scale_fac, atol=res*10,
+                    err_msg=f"Scale factor off by {s/scale_fac}")
+    xp_assert_close(d_s, desired_dual*scale_fac, atol=res*10, rtol=rel_tol_d,
+                    err_msg="Calculated incorrect scaled window!")
+
+    # check unscaled window:
+    d_u, u = closest_STFT_dual_window(win * scale_fac, nperseg - noverlap,
+                                      desired_dual, scaled=False)
+
+    # this should be hard-coded not computed, so no need for allclose
+    assert u == 1., "Scaling factor not 1 for parameter `scaled=True`!"
+    xp_assert_close(d_u, desired_dual, atol=res*10, rtol=rel_tol_d,
+                    err_msg="Calculated incorrect unscaled window!")
+
+
+def test_invalid_initializer_parameters():
+    """Verify that exceptions get raised on invalid parameters when
+    instantiating ShortTimeFFT. """
+    with pytest.raises(ValueError, match=r"Parameter win must be 1d, " +
+                                         r"but win.shape=\(2, 2\)!"):
+        ShortTimeFFT(np.ones((2, 2)), hop=4, fs=1)
+    with pytest.raises(ValueError, match="Parameter win must have " +
+                                         "finite entries"):
+        ShortTimeFFT(np.array([1, np.inf, 2, 3]), hop=4, fs=1)
+    with pytest.raises(ValueError, match="Parameter hop=0 is not " +
+                                         "an integer >= 1!"):
+        ShortTimeFFT(np.ones(4), hop=0, fs=1)
+    with pytest.raises(ValueError, match="Parameter hop=2.0 is not " +
+                                         "an integer >= 1!"):
+        # noinspection PyTypeChecker
+        ShortTimeFFT(np.ones(4), hop=2.0, fs=1)
+    with pytest.raises(ValueError, match=r"dual_win.shape=\(5,\) must equal " +
+                                         r"win.shape=\(4,\)!"):
+        ShortTimeFFT(np.ones(4), hop=2, fs=1, dual_win=np.ones(5))
+    with pytest.raises(ValueError, match="Parameter dual_win must be " +
+                                         "a finite array!"):
+        ShortTimeFFT(np.ones(3), hop=2, fs=1,
+                     dual_win=np.array([np.nan, 2, 3]))
+
+
+def test_exceptions_properties_methods():
+    """Verify that exceptions get raised when setting properties or calling
+    method of ShortTimeFFT to/with invalid values."""
+    SFT = ShortTimeFFT(np.ones(8), hop=4, fs=1)
+    with pytest.raises(ValueError, match="Sampling interval T=-1 must be " +
+                                         "positive!"):
+        SFT.T = -1
+    with pytest.raises(ValueError, match="Sampling frequency fs=-1 must be " +
+                                         "positive!"):
+        SFT.fs = -1
+    with pytest.raises(ValueError, match="fft_mode='invalid_typ' not in " +
+                                         r"\('twosided', 'centered', " +
+                                         r"'onesided', 'onesided2X'\)!"):
+        SFT.fft_mode = 'invalid_typ'
+    with pytest.raises(ValueError, match="For scaling is None, " +
+                                         "fft_mode='onesided2X' is invalid.*"):
+        SFT.fft_mode = 'onesided2X'
+    with pytest.raises(ValueError, match="Attribute mfft=7 needs to be " +
+                                         "at least the window length.*"):
+        SFT.mfft = 7
+    with pytest.raises(ValueError, match="scaling='invalid' not in.*"):
+        # noinspection PyTypeChecker
+        SFT.scale_to('invalid')
+    with pytest.raises(ValueError, match="phase_shift=3.0 has the unit .*"):
+        SFT.phase_shift = 3.0
+    with pytest.raises(ValueError, match="-mfft < phase_shift < mfft " +
+                                         "does not hold.*"):
+        SFT.phase_shift = 2*SFT.mfft
+    with pytest.raises(ValueError, match="Parameter padding='invalid' not.*"):
+        # noinspection PyTypeChecker
+        g = SFT._x_slices(np.zeros(16), k_off=0, p0=0, p1=1, padding='invalid')
+        next(g)  # execute generator
+    with pytest.raises(ValueError, match="Trend type must be 'linear' " +
+                                         "or 'constant'"):
+        # noinspection PyTypeChecker
+        SFT.stft_detrend(np.zeros(16), detr='invalid')
+    with pytest.raises(ValueError, match="Parameter detr=nan is not a str, " +
+                                         "function or None!"):
+        # noinspection PyTypeChecker
+        SFT.stft_detrend(np.zeros(16), detr=np.nan)
+    with pytest.raises(ValueError, match="Invalid Parameter p0=0, p1=200.*"):
+        SFT.p_range(100, 0, 200)
+
+    with pytest.raises(ValueError, match="f_axis=0 may not be equal to " +
+                                         "t_axis=0!"):
+        SFT.istft(np.zeros((SFT.f_pts, 2)), t_axis=0, f_axis=0)
+    with pytest.raises(ValueError, match=r"S.shape\[f_axis\]=2 must be equal" +
+                                         " to self.f_pts=5.*"):
+        SFT.istft(np.zeros((2, 2)))
+    with pytest.raises(ValueError, match=r"S.shape\[t_axis\]=1 needs to have" +
+                                         " at least 2 slices.*"):
+        SFT.istft(np.zeros((SFT.f_pts, 1)))
+    with pytest.raises(ValueError, match=r".*\(k1=100\) <= \(k_max=12\) " +
+                                         "is false!$"):
+        SFT.istft(np.zeros((SFT.f_pts, 3)), k1=100)
+    with pytest.raises(ValueError, match=r"\(k1=1\) - \(k0=0\) = 1 has to " +
+                                         "be at least.* length 4!"):
+        SFT.istft(np.zeros((SFT.f_pts, 3)), k0=0, k1=1)
+
+    with pytest.raises(ValueError, match=r"Parameter axes_seq='invalid' " +
+                                         r"not in \['tf', 'ft'\]!"):
+        # noinspection PyTypeChecker
+        SFT.extent(n=100, axes_seq='invalid')
+    with pytest.raises(ValueError, match="Attribute fft_mode=twosided must.*"):
+        SFT.fft_mode = 'twosided'
+        SFT.extent(n=100)
+
+
+@pytest.mark.parametrize('m', ('onesided', 'onesided2X'))
+def test_exceptions_fft_mode_complex_win(m: FFT_MODE_TYPE):
+    """Verify that one-sided spectra are not allowed with complex-valued
+    windows or with complex-valued signals.
+
+    The reason being, the `rfft` function only accepts real-valued input.
+    """
+    with pytest.raises(ValueError,
+                       match=f"One-sided spectra, i.e., fft_mode='{m}'.*"):
+        ShortTimeFFT(np.ones(8)*1j, hop=4, fs=1, fft_mode=m)
+
+    SFT = ShortTimeFFT(np.ones(8)*1j, hop=4, fs=1, fft_mode='twosided')
+    with pytest.raises(ValueError,
+                       match=f"One-sided spectra, i.e., fft_mode='{m}'.*"):
+        SFT.fft_mode = m
+
+    SFT = ShortTimeFFT(np.ones(8), hop=4, fs=1, scale_to='psd', fft_mode='onesided')
+    with pytest.raises(ValueError, match="Complex-valued `x` not allowed for self.*"):
+        SFT.stft(np.ones(8)*1j)
+    SFT.fft_mode = 'onesided2X'
+    with pytest.raises(ValueError, match="Complex-valued `x` not allowed for self.*"):
+        SFT.stft(np.ones(8)*1j)
+
+
+def test_invalid_fft_mode_RuntimeError():
+    """Ensure exception gets raised when property `fft_mode` is invalid. """
+    SFT = ShortTimeFFT(np.ones(8), hop=4, fs=1)
+    # noinspection PyTypeChecker
+    SFT._fft_mode = 'invalid_typ'
+
+    with pytest.raises(RuntimeError):
+        _ = SFT.f
+    with pytest.raises(RuntimeError):
+        SFT._fft_func(np.ones(8))
+    with pytest.raises(RuntimeError):
+        SFT._ifft_func(np.ones(8))
+
+
+@pytest.mark.parametrize('win_params, Nx', [(('gaussian', 2.), 9),  # in docstr
+                                            ('triang', 7),
+                                            (('kaiser', 4.0), 9),
+                                            (('exponential', None, 1.), 9),
+                                            (4.0, 9)])
+def test_from_window(win_params, Nx: int):
+    """Verify that `from_window()` handles parameters correctly.
+
+    The window parameterizations are documented in the `get_window` docstring.
+    """
+    w_sym, fs = get_window(win_params, Nx, fftbins=False), 16.
+    w_per = get_window(win_params, Nx, fftbins=True)
+    SFT0 = ShortTimeFFT(w_sym, hop=3, fs=fs, fft_mode='twosided',
+                        scale_to='psd', phase_shift=1)
+    nperseg = len(w_sym)
+    noverlap = nperseg - SFT0.hop
+    SFT1 = ShortTimeFFT.from_window(win_params, fs, nperseg, noverlap,
+                                    symmetric_win=True, fft_mode='twosided',
+                                    scale_to='psd', phase_shift=1)
+    # periodic window:
+    SFT2 = ShortTimeFFT.from_window(win_params, fs, nperseg, noverlap,
+                                    symmetric_win=False, fft_mode='twosided',
+                                    scale_to='psd', phase_shift=1)
+    # Be informative when comparing instances:
+    xp_assert_equal(SFT1.win, SFT0.win)
+    xp_assert_close(SFT2.win, w_per / np.sqrt(sum(w_per**2) * fs))
+    for n_ in ('hop', 'T', 'fft_mode', 'mfft', 'scaling', 'phase_shift'):
+        v0, v1, v2 = (getattr(SFT_, n_) for SFT_ in (SFT0, SFT1, SFT2))
+        assert v1 == v0, f"SFT1.{n_}={v1} does not equal SFT0.{n_}={v0}"
+        assert v2 == v0, f"SFT2.{n_}={v2} does not equal SFT0.{n_}={v0}"
+
+
+def test_from_win_equals_dual_exceptions():
+    """Raise all occurring exceptions in `ShortTimeFFT.from_closest_win_equals_dual`.
+    """
+    with pytest.raises(ValueError, match="Parameter desired_win is not 1d, but.*"):
+        ShortTimeFFT.from_win_equals_dual(np.ones((3, 4)), hop=1, fs=1)
+    with pytest.raises(ValueError, match="Parameter desired_win cannot be of int.*"):
+        ShortTimeFFT.from_win_equals_dual(np.ones(4, dtype=int), hop=1, fs=1)
+    with pytest.raises(ValueError, match="Parameter desired_win is not 1d, but.*"):
+        ShortTimeFFT.from_win_equals_dual(np.array([]), hop=1, fs=1)
+    with pytest.raises(ValueError, match="Parameter desired_win must have finite.*"):
+        ShortTimeFFT.from_win_equals_dual(np.ones(3) * np.inf, hop=1, fs=1)
+    with pytest.raises(ValueError, match="Parameter hop=0 is not an integer .*"):
+        ShortTimeFFT.from_win_equals_dual(np.ones(4), hop=0, fs=1)
+    with pytest.raises(ValueError, match="Parameter hop=5 is not an integer .*"):
+        ShortTimeFFT.from_win_equals_dual(np.ones(4), hop=5, fs=1)
+
+    with pytest.raises(ValueError, match="P.+ desired_win does not have valid.*"):
+        w = np.array([1, 0, 1, 0, 1], dtype=float)
+        ShortTimeFFT.from_win_equals_dual(w, hop=2, fs=1)
+    with pytest.raises(ValueError, match="Parameter scale_to='invalid' not in.*"):
+        # noinspection PyTypeChecker
+        ShortTimeFFT.from_win_equals_dual(w, hop=2, fs=1, scale_to='invalid')
+
+@pytest.mark.parametrize('fft_bins', (True, False))
+@pytest.mark.parametrize('m, hop', [(16, 8), (16, 7), (17, 8), (17, 9), (16, 16)])
+def test_from_win_equals_dual_params(m, hop, fft_bins):
+    """Test windows parameterizations for `ShortTimeFFT.from_closest_win_equals_dual`.
+
+    The flattop window is used since it also has negative values.
+    """
+    desired_win = get_window('flattop', m, fftbins=fft_bins)
+    SFT0 = ShortTimeFFT.from_win_equals_dual(desired_win, hop, fs=1)
+    xp_assert_close(SFT0.dual_win, SFT0.win, err_msg="win must equals dual window!")
+
+    SFT1 = ShortTimeFFT(SFT0.win, hop, fs=1)
+    xp_assert_close(SFT1.dual_win, SFT0.win, err_msg="dual win isn't canonical win!")
+
+
+@pytest.mark.parametrize('fft_bins', (True, False))
+@pytest.mark.parametrize('m, hop', [(16, 8), (16, 7), (17, 8), (17, 9), (16, 16)])
+@pytest.mark.parametrize('scale_to', (None, 'unitary'))
+def test_from_win_equals_dual_roundtrip(m, hop, fft_bins, scale_to):
+    """Testing roundtrip verifies that the dual window is correct.
+    """
+    desired_win = get_window('flattop', m, fftbins=fft_bins)
+    SFT0 = ShortTimeFFT.from_win_equals_dual(desired_win, hop, fs=1)
+
+    x = np.cos(np.arange(2*m)**2)
+    x1 = SFT0.istft(SFT0.stft(x), 0, len(x))
+    xp_assert_close(x1, x, err_msg="Roundtrip for win equaling its dual failed!")
+
+
+def test_from_win_equals_dual_unitary():
+    """Check that STFT can be unitary mapping. """
+    m, hop = 8, 4
+    des_win = get_window('hann', m)
+    SFT = ShortTimeFFT.from_win_equals_dual(des_win, hop, 1, fft_mode='twosided',
+                                            scale_to='unitary')
+    # Orthogonal signals:
+    x, y = np.tile([-1, -1, 1, 1], 4), np.tile([1, -1, -1, 1], 4)
+    Sxx, Sxy = SFT.spectrogram(x), SFT.spectrogram(x, y)
+
+    atol = np.finfo(Sxx.dtype).resolution
+    assert sum(x * y) == 0
+    xp_assert_close(np.sum(Sxx), np.sum(x ** 2, dtype=Sxx.dtype), atol=atol,
+                    err_msg="Energies do not match!")
+    xp_assert_close(np.sum(Sxy), 0.0j, atol=atol,
+                    err_msg="STFT scalar product of Sx and Sy not 0!")
+    xp_assert_close(SFT.dual_win, SFT.win*SFT.m_num, atol=atol,
+                    err_msg="Wrong factor for dual_win/win!")
+
+
+def test_dual_win_roundtrip():
+    """Verify the duality of `win` and `dual_win`.
+
+    Note that this test does not work for arbitrary windows, since dual windows
+    are not unique. It always works for invertible STFTs if the windows do not
+    overlap.
+    """
+    # Non-standard values for keyword arguments (except for `scale_to`):
+    kw = dict(hop=4, fs=1, fft_mode='twosided', mfft=8, scale_to=None,
+              phase_shift=2)
+    SFT0 = ShortTimeFFT(np.ones(4), **kw)
+    SFT1 = ShortTimeFFT.from_dual(SFT0.dual_win, **kw)
+    xp_assert_close(SFT1.dual_win, SFT0.win)
+
+
+@pytest.mark.parametrize('scale_to, fac_psd, fac_mag',
+                         [(None, 0.25, 0.125),
+                          ('magnitude', 2.0, 1),
+                          ('psd', 1, 0.5)])
+def test_scaling(scale_to: Literal['magnitude', 'psd'], fac_psd, fac_mag):
+    """Verify scaling calculations.
+
+    * Verify passing `scale_to`parameter  to ``__init__().
+    * Roundtrip while changing scaling factor.
+    """
+    SFT = ShortTimeFFT(np.ones(4) * 2, hop=4, fs=1, scale_to=scale_to)
+    assert SFT.fac_psd == fac_psd
+    assert SFT.fac_magnitude == fac_mag
+    # increase coverage by accessing properties twice:
+    assert SFT.fac_psd == fac_psd
+    assert SFT.fac_magnitude == fac_mag
+
+    x = np.fft.irfft([0, 0, 7, 0, 0, 0, 0])  # periodic signal
+    Sx = SFT.stft(x)
+    Sx_mag, Sx_psd = Sx * SFT.fac_magnitude, Sx * SFT.fac_psd
+
+    SFT.scale_to('magnitude')
+    x_mag = SFT.istft(Sx_mag, k1=len(x))
+    xp_assert_close(x_mag, x)
+
+    SFT.scale_to('psd')
+    x_psd = SFT.istft(Sx_psd, k1=len(x))
+    xp_assert_close(x_psd, x)
+
+
+def test_scale_to():
+    """Verify `scale_to()` method."""
+    SFT = ShortTimeFFT(np.ones(4) * 2, hop=4, fs=1, scale_to=None)
+
+    SFT.scale_to('magnitude')
+    assert SFT.scaling == 'magnitude'
+    assert SFT.fac_psd == 2.0
+    assert SFT.fac_magnitude == 1
+
+    SFT.scale_to('psd')
+    assert SFT.scaling == 'psd'
+    assert SFT.fac_psd == 1
+    assert SFT.fac_magnitude == 0.5
+
+    SFT.scale_to('psd')  # needed for coverage
+
+    for scale, s_fac in zip(('magnitude', 'psd'), (8, 4)):
+        SFT = ShortTimeFFT(np.ones(4) * 2, hop=4, fs=1, scale_to=None)
+        dual_win = SFT.dual_win.copy()
+
+        SFT.scale_to(cast(Literal['magnitude', 'psd'], scale))
+        xp_assert_close(SFT.dual_win, dual_win * s_fac)
+
+
+def test_x_slices_padding():
+    """Verify padding.
+
+    The reference arrays were taken from  the docstrings of `zero_ext`,
+    `const_ext`, `odd_ext()`, and `even_ext()` from the _array_tools module.
+    """
+    SFT = ShortTimeFFT(np.ones(5), hop=4, fs=1)
+    x = np.array([[1, 2, 3, 4, 5], [0, 1, 4, 9, 16]], dtype=float)
+    d = {'zeros': [[[0, 0, 1, 2, 3], [0, 0, 0, 1, 4]],
+                   [[3, 4, 5, 0, 0], [4, 9, 16, 0, 0]]],
+         'edge': [[[1, 1, 1, 2, 3], [0, 0, 0, 1, 4]],
+                  [[3, 4, 5, 5, 5], [4, 9, 16, 16, 16]]],
+         'even': [[[3, 2, 1, 2, 3], [4, 1, 0, 1, 4]],
+                  [[3, 4, 5, 4, 3], [4, 9, 16, 9, 4]]],
+         'odd': [[[-1, 0, 1, 2, 3], [-4, -1, 0, 1, 4]],
+                 [[3, 4, 5, 6, 7], [4, 9, 16, 23, 28]]]}
+    for p_, xx in d.items():
+        gen = SFT._x_slices(np.array(x), 0, 0, 2, padding=cast(PAD_TYPE, p_))
+        yy = np.array([y_.copy() for y_ in gen])  # due to inplace copying
+        xx = np.asarray(xx, dtype=np.float64)
+        xp_assert_equal(yy, xx, err_msg=f"Failed '{p_}' padding.")
+
+
+def test_invertible():
+    """Verify `invertible` property. """
+    SFT = ShortTimeFFT(np.ones(8), hop=4, fs=1)
+    assert SFT.invertible
+    SFT = ShortTimeFFT(np.ones(8), hop=9, fs=1)
+    assert not SFT.invertible
+
+
+def test_border_values():
+    """Ensure that minimum and maximum values of slices are correct."""
+    SFT = ShortTimeFFT(np.ones(8), hop=4, fs=1)
+    assert SFT.p_min == 0
+    assert SFT.k_min == -4
+    assert SFT.lower_border_end == (4, 1)
+    assert SFT.lower_border_end == (4, 1)  # needed to test caching
+    assert SFT.p_max(10) == 4
+    assert SFT.k_max(10) == 16
+    assert SFT.upper_border_begin(10) == (4, 2)
+    # Raise exceptions:
+    with pytest.raises(ValueError, match="^Parameter n must be"):
+        SFT.upper_border_begin(3)
+    with pytest.raises(ValueError, match="^Parameter n must be"):
+        SFT._post_padding(3)
+
+def test_border_values_exotic():
+    """Ensure that the border calculations are correct for windows with
+    zeros. """
+    w = np.array([0, 0, 0, 0, 0, 0, 0, 1.])
+    SFT = ShortTimeFFT(w, hop=1, fs=1)
+    assert SFT.lower_border_end == (0, 0)
+
+    SFT = ShortTimeFFT(np.flip(w), hop=20, fs=1)
+    assert SFT.upper_border_begin(4) == (16, 1)
+    assert SFT.upper_border_begin(5) == (16, 1)
+    assert SFT.upper_border_begin(23) == (36, 2)
+    assert SFT.upper_border_begin(24) == (36, 2)
+    assert SFT.upper_border_begin(25) == (36, 2)
+
+    SFT._hop = -1  # provoke unreachable line
+    with pytest.raises(RuntimeError):
+        _ = SFT.k_max(4)
+    with pytest.raises(RuntimeError):
+        _ = SFT.k_min
+
+
+def test_t():
+    """Verify that the times of the slices are correct. """
+    SFT = ShortTimeFFT(np.ones(8), hop=4, fs=2)
+    assert SFT.T == 1/2
+    assert SFT.fs == 2.
+    assert SFT.delta_t == 4 * 1/2
+    t_stft = np.arange(0, SFT.p_max(10)) * SFT.delta_t
+    xp_assert_equal(SFT.t(10), t_stft)
+    xp_assert_equal(SFT.t(10, 1, 3), t_stft[1:3])
+    SFT.T = 1/4
+    assert SFT.T == 1/4
+    assert SFT.fs == 4
+    SFT.fs = 1/8
+    assert SFT.fs == 1/8
+    assert SFT.T == 8
+
+
+@pytest.mark.parametrize('fft_mode, f',
+                         [('onesided', [0., 1., 2.]),
+                          ('onesided2X', [0., 1., 2.]),
+                          ('twosided', [0., 1., 2., -2., -1.]),
+                          ('centered', [-2., -1., 0., 1., 2.])])
+def test_f(fft_mode: FFT_MODE_TYPE, f):
+    """Verify the frequency values property `f`."""
+    SFT = ShortTimeFFT(np.ones(5), hop=4, fs=5, fft_mode=fft_mode,
+                       scale_to='psd')
+    xp_assert_equal(SFT.f, f)
+
+
+@pytest.mark.parametrize('n', [20, 21])
+@pytest.mark.parametrize('m', [5, 6])
+@pytest.mark.parametrize('fft_mode', ['onesided', 'centered'])
+def test_extent(n, m, fft_mode: FFT_MODE_TYPE):
+    """Ensure that the `extent()` method is correct. """
+    SFT = ShortTimeFFT(np.ones(m), hop=m, fs=m, fft_mode=fft_mode)
+
+    t0 = SFT.t(n)[0]  # first timestamp
+    t1 = SFT.t(n)[-1] + SFT.delta_t  # last timestamp + 1
+    t0c, t1c = t0 - SFT.delta_t / 2, t1 - SFT.delta_t / 2  # centered timestamps
+
+    f0 = SFT.f[0]  # first frequency
+    f1 = SFT.f[-1] + SFT.delta_f  # last frequency + 1
+    f0c, f1c = f0 - SFT.delta_f / 2, f1 - SFT.delta_f / 2  # centered frequencies
+
+    assert SFT.extent(n, 'tf', False) == (t0, t1, f0, f1)
+    assert SFT.extent(n, 'ft', False) == (f0, f1, t0, t1)
+    assert SFT.extent(n, 'tf', True) == (t0c, t1c, f0c, f1c)
+    assert SFT.extent(n, 'ft', True) == (f0c, f1c, t0c, t1c)
+
+
+def test_spectrogram():
+    """Verify spectrogram and cross-spectrogram methods. """
+    SFT = ShortTimeFFT(np.ones(8), hop=4, fs=1)
+    x, y = np.ones(10), np.arange(10)
+    X, Y = SFT.stft(x), SFT.stft(y)
+    xp_assert_close(SFT.spectrogram(x), X.real**2+X.imag**2)
+    xp_assert_close(SFT.spectrogram(x, y), X * Y.conj())
+
+
+@pytest.mark.parametrize('n', [8, 9])
+def test_fft_func_roundtrip(n: int):
+    """Test roundtrip `ifft_func(fft_func(x)) == x` for all permutations of
+    relevant parameters. """
+    np.random.seed(2394795)
+    x0 = np.random.rand(n)
+    w, h_n = np.ones(n), 4
+
+    pp = dict(
+        fft_mode=get_args(FFT_MODE_TYPE),
+        mfft=[None, n, n+1, n+2],
+        scaling=[None, 'magnitude', 'psd'],
+        phase_shift=[None, -n+1, 0, n // 2, n-1])
+    for f_typ, mfft, scaling, phase_shift in product(*pp.values()):
+        if f_typ == 'onesided2X' and scaling is None:
+            continue  # this combination is forbidden
+        SFT = ShortTimeFFT(w, h_n, fs=n, fft_mode=f_typ, mfft=mfft,
+                           scale_to=scaling, phase_shift=phase_shift)
+        X0 = SFT._fft_func(x0)
+        x1 = SFT._ifft_func(X0)
+        xp_assert_close(x0.astype(x1.dtype), x1,
+                        err_msg="_fft_func() roundtrip failed for " +
+                        f"{f_typ=}, {mfft=}, {scaling=}, {phase_shift=}")
+
+    SFT = ShortTimeFFT(w, h_n, fs=1)
+    SFT._fft_mode = 'invalid_fft'  # type: ignore
+    with pytest.raises(RuntimeError):
+        SFT._fft_func(x0)
+    with pytest.raises(RuntimeError):
+        SFT._ifft_func(x0)
+
+
+@pytest.mark.parametrize('i', range(19))
+def test_impulse_roundtrip(i):
+    """Roundtrip for an impulse being at different positions `i`."""
+    n = 19
+    w, h_n = np.ones(8), 3
+    x = np.zeros(n)
+    x[i] = 1
+
+    SFT = ShortTimeFFT(w, hop=h_n, fs=1, scale_to=None, phase_shift=None)
+    Sx = SFT.stft(x)
+    # test slicing the input signal into two parts:
+    n_q = SFT.nearest_k_p(n // 2)
+    Sx0 = SFT.stft(x[:n_q], padding='zeros')
+    Sx1 = SFT.stft(x[n_q:], padding='zeros')
+    q0_ub = SFT.upper_border_begin(n_q)[1] - SFT.p_min
+    q1_le = SFT.lower_border_end[1] - SFT.p_min
+    xp_assert_close(Sx0[:, :q0_ub], Sx[:, :q0_ub], err_msg=f"{i=}")
+    xp_assert_close(Sx1[:, q1_le:], Sx[:, q1_le-Sx1.shape[1]:],
+                    err_msg=f"{i=}")
+
+    Sx01 = np.hstack((Sx0[:, :q0_ub],
+                      Sx0[:, q0_ub:] + Sx1[:, :q1_le],
+                      Sx1[:, q1_le:]))
+    xp_assert_close(Sx, Sx01, atol=1e-8, err_msg=f"{i=}")
+
+    y = SFT.istft(Sx, 0, n)
+    xp_assert_close(y, x, atol=1e-8, err_msg=f"{i=}")
+    y0 = SFT.istft(Sx, 0, n//2)
+    xp_assert_close(x[:n//2], y0, atol=1e-8, err_msg=f"{i=}")
+    y1 = SFT.istft(Sx, n // 2, n)
+    xp_assert_close(x[n // 2:], y1, atol=1e-8, err_msg=f"{i=}")
+
+
+@pytest.mark.parametrize('hop', [1, 7, 8])
+def test_asymmetric_window_roundtrip(hop: int):
+    """An asymmetric window could uncover indexing problems. """
+    np.random.seed(23371)
+
+    w = np.arange(16) / 8  # must be of type float
+    w[len(w)//2:] = 1
+    SFT = ShortTimeFFT(w, hop, fs=1)
+
+    x = 10 * np.random.randn(64)
+    Sx = SFT.stft(x)
+    x1 = SFT.istft(Sx, k1=len(x))
+    xp_assert_close(x1, x1, err_msg="Roundtrip for asymmetric window with " +
+                                    f" {hop=} failed!")
+
+
+@pytest.mark.parametrize('m_num', [6, 7])
+def test_minimal_length_signal(m_num):
+    """Verify that the shortest allowed signal works. """
+    SFT = ShortTimeFFT(np.ones(m_num), m_num//2, fs=1)
+    n = math.ceil(m_num/2)
+    x = np.ones(n)
+    Sx = SFT.stft(x)
+    x1 = SFT.istft(Sx, k1=n)
+    xp_assert_close(x1, x, err_msg=f"Roundtrip minimal length signal ({n=})" +
+                                   f" for {m_num} sample window failed!")
+    with pytest.raises(ValueError, match=rf"len\(x\)={n-1} must be >= ceil.*"):
+        SFT.stft(x[:-1])
+    with pytest.raises(ValueError, match=rf"S.shape\[t_axis\]={Sx.shape[1]-1}"
+                       f" needs to have at least {Sx.shape[1]} slices"):
+        SFT.istft(Sx[:, :-1], k1=n)
+
+
+def test_compare_stft_detrend():
+    """Test the detrending in `ShortTimeFFT.stft_detrend()`. """
+    SFT = ShortTimeFFT(np.ones(4), 4, fs=1)
+    x0 = np.zeros(4) # signal without trend
+    x1 = x0 + 3  # signal with constant trend
+    x2 = x0 + np.arange(len(x0))  # signal with linear trend
+
+    kw = dict(k_offset=2, p1=1)  # we want only one slice
+    Sx0 = SFT.stft(x0, **kw)  # no trend
+    Sx1 = SFT.stft_detrend(x1, detr='constant', **kw)
+    Sx2 = SFT.stft_detrend(x2, detr='linear', **kw)
+    Sx3 = SFT.stft_detrend(x1, detr=lambda x: x - np.mean(x), **kw)
+
+    atol = np.finfo(Sx0.dtype).resolution * 5  # needed to compare with array of zeros
+    xp_assert_close(Sx1, Sx0, atol=atol, err_msg="Constant detrending failed!")
+    xp_assert_close(Sx0, Sx2, atol=atol, err_msg="Linear detrending failed!")
+    xp_assert_close(Sx0, Sx3, atol=atol, err_msg="Detrending using a function failed!")
+
+
+def test_tutorial_stft_sliding_win():
+    """Verify example in "Sliding Windows" subsection from the "User Guide".
+
+    In :ref:`tutorial_stft_sliding_win` (file ``signal.rst``) of the
+    :ref:`user_guide` the behavior the border behavior of
+    ``ShortTimeFFT(np.ones(6), 2, fs=1)`` with a 50 sample signal is discussed.
+    This test verifies the presented indexes.
+    """
+    SFT = ShortTimeFFT(np.ones(6), 2, fs=1)
+
+    # Lower border:
+    assert SFT.m_num_mid == 3, f"Slice middle is not 3 but {SFT.m_num_mid=}"
+    assert SFT.p_min == -1, f"Lowest slice {SFT.p_min=} is not -1"
+    assert SFT.k_min == -5, f"Lowest slice sample {SFT.p_min=} is not -5"
+    k_lb, p_lb = SFT.lower_border_end
+    assert p_lb == 2, f"First unaffected slice {p_lb=} is not 2"
+    assert k_lb == 5, f"First unaffected sample {k_lb=} is not 5"
+
+    n = 50  # upper signal border
+    assert (p_max := SFT.p_max(n)) == 27, f"Last slice {p_max=} must be 27"
+    assert (k_max := SFT.k_max(n)) == 55, f"Last sample {k_max=} must be 55"
+    k_ub, p_ub = SFT.upper_border_begin(n)
+    assert p_ub == 24, f"First upper border slice {p_ub=} must be 24"
+    assert k_ub == 45, f"First upper border slice {k_ub=} must be 45"
+
+
+def test_tutorial_stft_legacy_stft():
+    """Verify STFT example in "Comparison with Legacy Implementation" from the
+    "User Guide".
+
+    In :ref:`tutorial_stft_legacy_stft` (file ``signal.rst``) of the
+    :ref:`user_guide` the legacy and the new implementation are compared.
+    """
+    fs, N = 200, 1001  # # 200 Hz sampling rate for 5 s signal
+    t_z = np.arange(N) / fs  # time indexes for signal
+    z = np.exp(2j*np.pi * 70 * (t_z - 0.2 * t_z ** 2))  # complex-valued chirp
+
+    nperseg, noverlap = 50, 40
+    win = ('gaussian', 1e-2 * fs)  # Gaussian with 0.01 s standard deviation
+
+    # Legacy STFT:
+    f0_u, t0, Sz0_u = stft(z, fs, win, nperseg, noverlap,
+                           return_onesided=False, scaling='spectrum')
+    Sz0 = fftshift(Sz0_u, axes=0)
+
+    # New STFT:
+    SFT = ShortTimeFFT.from_window(win, fs, nperseg, noverlap,
+                                   fft_mode='centered',
+                                   scale_to='magnitude', phase_shift=None)
+    Sz1 = SFT.stft(z)
+
+    xp_assert_close(Sz0, Sz1[:, 2:-1])
+
+    xp_assert_close((abs(Sz1[:, 1]).min(), abs(Sz1[:, 1]).max()),
+                    (6.925060911593139e-07, 8.00271269218721e-07))
+
+    t0_r, z0_r = istft(Sz0_u, fs, win, nperseg, noverlap, input_onesided=False,
+                       scaling='spectrum')
+    z1_r = SFT.istft(Sz1, k1=N)
+    assert len(z0_r) == N + 9
+    xp_assert_close(z0_r[:N], z)
+    xp_assert_close(z1_r, z)
+
+    #  Spectrogram is just the absolute square of th STFT:
+    xp_assert_close(SFT.spectrogram(z), abs(Sz1) ** 2)
+
+
+def test_tutorial_stft_legacy_spectrogram():
+    """Verify spectrogram example in "Comparison with Legacy Implementation"
+    from the "User Guide".
+
+    In :ref:`tutorial_stft_legacy_stft` (file ``signal.rst``) of the
+    :ref:`user_guide` the legacy and the new implementation are compared.
+    """
+    fs, N = 200, 1001  # 200 Hz sampling rate for almost 5 s signal
+    t_z = np.arange(N) / fs  # time indexes for signal
+    z = np.exp(2j*np.pi*70 * (t_z - 0.2*t_z**2))  # complex-valued sweep
+
+    nperseg, noverlap = 50, 40
+    win = ('gaussian', 1e-2 * fs)  # Gaussian with 0.01 s standard dev.
+
+    # Legacy spectrogram:
+    f2_u, t2, Sz2_u = spectrogram(z, fs, win, nperseg, noverlap, detrend=None,
+                                  return_onesided=False, scaling='spectrum',
+                                  mode='complex')
+
+    f2, Sz2 = fftshift(f2_u), fftshift(Sz2_u, axes=0)
+
+    # New STFT:
+    SFT = ShortTimeFFT.from_window(win, fs, nperseg, noverlap,
+                                   fft_mode='centered', scale_to='magnitude',
+                                   phase_shift=None)
+    Sz3 = SFT.stft(z, p0=0, p1=(N-noverlap) // SFT.hop, k_offset=nperseg // 2)
+    t3 = SFT.t(N, p0=0, p1=(N-noverlap) // SFT.hop, k_offset=nperseg // 2)
+
+    xp_assert_close(t2, t3)
+    xp_assert_close(f2, SFT.f)
+    xp_assert_close(Sz2, Sz3)
+
+
+def test_permute_axes():
+    """Verify correctness of four-dimensional signal by permuting its
+    shape. """
+    n = 25
+    SFT = ShortTimeFFT(np.ones(8)/8, hop=3, fs=n)
+    x0 = np.arange(n, dtype=np.float64)
+    Sx0 = SFT.stft(x0)
+    Sx0 = Sx0.reshape((Sx0.shape[0], 1, 1, 1, Sx0.shape[-1]))
+    SxT = np.moveaxis(Sx0, (0, -1), (-1, 0))
+
+    atol = 2 * np.finfo(SFT.win.dtype).resolution
+    for i in range(4):
+        y = np.reshape(x0, np.roll((n, 1, 1, 1), i))
+        Sy = SFT.stft(y, axis=i)
+        xp_assert_close(Sy, np.moveaxis(Sx0, 0, i))
+
+        yb0 = SFT.istft(Sy, k1=n, f_axis=i)
+        xp_assert_close(yb0, y, atol=atol)
+        # explicit t-axis parameter (for coverage):
+        yb1 = SFT.istft(Sy, k1=n, f_axis=i, t_axis=Sy.ndim-1)
+        xp_assert_close(yb1, y, atol=atol)
+
+        SyT = np.moveaxis(Sy, (i, -1), (-1, i))
+        xp_assert_close(SyT, np.moveaxis(SxT, 0, i))
+
+        ybT = SFT.istft(SyT, k1=n, t_axis=i, f_axis=-1)
+        xp_assert_close(ybT, y, atol=atol)
+
+
+@pytest.mark.parametrize("fft_mode",
+                         ('twosided', 'centered', 'onesided', 'onesided2X'))
+def test_roundtrip_multidimensional(fft_mode: FFT_MODE_TYPE):
+    """Test roundtrip of a multidimensional input signal versus its components.
+
+    This test can uncover potential problems with `fftshift()`.
+    """
+    n = 9
+    x = np.arange(4*n*2, dtype=np.float64).reshape(4, n, 2)
+    SFT = ShortTimeFFT(get_window('hann', 4), hop=2, fs=1,
+                       scale_to='magnitude', fft_mode=fft_mode)
+    Sx = SFT.stft(x, axis=1)
+    y = SFT.istft(Sx, k1=n, f_axis=1, t_axis=-1)
+    xp_assert_close(y, x.astype(y.dtype), err_msg='Multidim. roundtrip failed!')
+
+    for i, j in product(range(x.shape[0]), range(x.shape[2])):
+        y_ = SFT.istft(Sx[i, :, j, :], k1=n)
+        xp_assert_close(y_, x[i, :, j].astype(y_.dtype),
+                        err_msg="Multidim. roundtrip for component " +
+                        f"x[{i}, :, {j}] and {fft_mode=} failed!")
+
+@pytest.mark.parametrize("phase_shift", (0, 4,  None))
+def test_roundtrip_two_dimensional(phase_shift: int|None):
+    """Test roundtrip of a 2 channel input signal with `mfft` set with different
+    values for `phase_shift`
+
+    Tests for Issue https://github.com/scipy/scipy/issues/21671
+    """
+    n = 21
+    SFT = ShortTimeFFT.from_window('hann', fs=1, nperseg=13, noverlap=7,
+                                   mfft=16, phase_shift=phase_shift)
+    x = np.arange(2*n, dtype=float).reshape(2, n)
+    Sx = SFT.stft(x)
+    y = SFT.istft(Sx, k1=n)
+    xp_assert_close(y, x, atol=2 * np.finfo(SFT.win.dtype).resolution,
+                    err_msg='2-dim. roundtrip failed!')
+
+
+@pytest.mark.parametrize('window, n, nperseg, noverlap',
+                         [('boxcar', 100, 10, 0),     # Test no overlap
+                          ('boxcar', 100, 10, 9),     # Test high overlap
+                          ('bartlett', 101, 51, 26),  # Test odd nperseg
+                          ('hann', 1024, 256, 128),   # Test defaults
+                          (('tukey', 0.5), 1152, 256, 64),  # Test Tukey
+                          ('hann', 1024, 256, 255),   # Test overlapped hann
+                          ('boxcar', 100, 10, 3),     # NOLA True, COLA False
+                          ('bartlett', 101, 51, 37),  # NOLA True, COLA False
+                          ('hann', 1024, 256, 127),   # NOLA True, COLA False
+                          # NOLA True, COLA False:
+                          (('tukey', 0.5), 1152, 256, 14),
+                          ('hann', 1024, 256, 5)])    # NOLA True, COLA False
+def test_roundtrip_windows(window, n: int, nperseg: int, noverlap: int):
+    """Roundtrip test adapted from `test_spectral.TestSTFT`.
+
+    The parameters are taken from the methods test_roundtrip_real(),
+    test_roundtrip_nola_not_cola(), test_roundtrip_float32(),
+    test_roundtrip_complex().
+    """
+    np.random.seed(2394655)
+
+    w = get_window(window, nperseg)
+    SFT = ShortTimeFFT(w, nperseg - noverlap, fs=1, fft_mode='twosided',
+                       phase_shift=None)
+
+    z = 10 * np.random.randn(n) + 10j * np.random.randn(n)
+    Sz = SFT.stft(z)
+    z1 = SFT.istft(Sz, k1=len(z))
+    xp_assert_close(z, z1, err_msg="Roundtrip for complex values failed")
+
+    x = 10 * np.random.randn(n)
+    Sx = SFT.stft(x)
+    x1 = SFT.istft(Sx, k1=len(z))
+    xp_assert_close(x.astype(np.complex128), x1,
+                    err_msg="Roundtrip for float values failed")
+
+    x32 = x.astype(np.float32)
+    Sx32 = SFT.stft(x32)
+    x32_1 = SFT.istft(Sx32, k1=len(x32))
+    x32_1_r = x32_1.real
+    xp_assert_close(x32, x32_1_r.astype(np.float32),
+                    err_msg="Roundtrip for 32 Bit float values failed")
+    xp_assert_close(x32.imag, np.zeros_like(x32.imag),
+                    err_msg="Roundtrip for 32 Bit float values failed")
+
+
+@pytest.mark.parametrize('signal_type', ('real', 'complex'))
+def test_roundtrip_complex_window(signal_type):
+    """Test roundtrip for complex-valued window function
+
+    The purpose of this test is to check if the dual window is calculated
+    correctly for complex-valued windows.
+    """
+    np.random.seed(1354654)
+    win = np.exp(2j*np.linspace(0, np.pi, 8))
+    SFT = ShortTimeFFT(win, 3, fs=1, fft_mode='twosided')
+
+    z = 10 * np.random.randn(11)
+    if signal_type == 'complex':
+        z = z + 2j * z
+    Sz = SFT.stft(z)
+    z1 = SFT.istft(Sz, k1=len(z))
+    xp_assert_close(z.astype(np.complex128), z1,
+                    err_msg="Roundtrip for complex-valued window failed")
+
+
+def test_average_all_segments():
+    """Compare `welch` function with stft mean.
+
+    Ported from `TestSpectrogram.test_average_all_segments` from file
+    ``test__spectral.py``.
+    """
+    x = np.random.randn(1024)
+
+    fs = 1.0
+    window = ('tukey', 0.25)
+    nperseg, noverlap = 16, 2
+    fw, Pw = welch(x, fs, window, nperseg, noverlap)
+    SFT = ShortTimeFFT.from_window(window, fs, nperseg, noverlap,
+                                   fft_mode='onesided2X', scale_to='psd',
+                                   phase_shift=None)
+    # `welch` positions the window differently than the STFT:
+    P = SFT.spectrogram(x, detr='constant', p0=0,
+                        p1=(len(x)-noverlap)//SFT.hop, k_offset=nperseg//2)
+
+    xp_assert_close(SFT.f, fw)
+    xp_assert_close(np.mean(P, axis=-1), Pw)
+
+
+@pytest.mark.parametrize('window, N, nperseg, noverlap, mfft',
+                         # from test_roundtrip_padded_FFT:
+                         [('hann', 1024, 256, 128, 512),
+                          ('hann', 1024, 256, 128, 501),
+                          ('boxcar', 100, 10, 0, 33),
+                          (('tukey', 0.5), 1152, 256, 64, 1024),
+                          # from test_roundtrip_padded_signal:
+                          ('boxcar', 101, 10, 0, None),
+                          ('hann', 1000, 256, 128, None),
+                          # from test_roundtrip_boundary_extension:
+                          ('boxcar', 100, 10, 0, None),
+                          ('boxcar', 100, 10, 9, None)])
+@pytest.mark.parametrize('padding', get_args(PAD_TYPE))
+def test_stft_padding_roundtrip(window, N: int, nperseg: int, noverlap: int,
+                                mfft: int, padding):
+    """Test the parameter 'padding' of `stft` with roundtrips.
+
+    The STFT parametrizations were taken from the methods
+    `test_roundtrip_padded_FFT`, `test_roundtrip_padded_signal` and
+    `test_roundtrip_boundary_extension` from class `TestSTFT` in  file
+    ``test_spectral.py``. Note that the ShortTimeFFT does not need the
+    concept of "boundary extension".
+    """
+    x = normal_distribution.rvs(size=N, random_state=2909)  # real signal
+    z = x * np.exp(1j * np.pi / 4)  # complex signal
+
+    SFT = ShortTimeFFT.from_window(window, 1, nperseg, noverlap,
+                                   fft_mode='twosided', mfft=mfft)
+    Sx = SFT.stft(x, padding=padding)
+    x1 = SFT.istft(Sx, k1=N)
+    xp_assert_close(x1, x.astype(np.complex128),
+                    err_msg=f"Failed real roundtrip with '{padding}' padding")
+
+    Sz = SFT.stft(z, padding=padding)
+    z1 = SFT.istft(Sz, k1=N)
+    xp_assert_close(z1, z, err_msg="Failed complex roundtrip with " +
+                    f" '{padding}' padding")
+
+
+@pytest.mark.parametrize('N_x', (128, 129, 255, 256, 1337))  # signal length
+@pytest.mark.parametrize('w_size', (128, 256))  # window length
+@pytest.mark.parametrize('t_step', (4, 64))  # SFT time hop
+@pytest.mark.parametrize('f_c', (7., 23.))  # frequency of input sine
+def test_energy_conservation(N_x: int, w_size: int, t_step: int, f_c: float):
+    """Test if a `psd`-scaled STFT conserves the L2 norm.
+
+    This test is adapted from MNE-Python [1]_. Besides being battle-tested,
+    this test has the benefit of using non-standard window including
+    non-positive values and a 2d input signal.
+
+    Since `ShortTimeFFT` requires the signal length `N_x` to be at least the
+    window length `w_size`, the parameter `N_x` was changed from
+    ``(127, 128, 255, 256, 1337)`` to ``(128, 129, 255, 256, 1337)`` to be
+    more useful.
+
+    .. [1] File ``test_stft.py`` of MNE-Python
+        https://github.com/mne-tools/mne-python/blob/main/mne/time_frequency/tests/test_stft.py
+    """
+    window = np.sin(np.arange(.5, w_size + .5) / w_size * np.pi)
+    SFT = ShortTimeFFT(window, t_step, fs=1000, fft_mode='onesided2X',
+                       scale_to='psd')
+    atol = 2*np.finfo(window.dtype).resolution
+    N_x = max(N_x, w_size)  # minimal sing
+    # Test with low frequency signal
+    t = np.arange(N_x).astype(np.float64)
+    x = np.sin(2 * np.pi * f_c * t * SFT.T)
+    x = np.array([x, x + 1.])
+    X = SFT.stft(x)
+    xp = SFT.istft(X, k1=N_x)
+
+    max_freq = SFT.f[np.argmax(np.sum(np.abs(X[0]) ** 2, axis=1))]
+
+    assert X.shape[1] == SFT.f_pts
+    assert np.all(SFT.f >= 0.)
+    assert np.abs(max_freq - f_c) < 1.
+    xp_assert_close(x, xp, atol=atol)
+
+    # check L2-norm squared (i.e., energy) conservation:
+    E_x = np.sum(x**2, axis=-1) * SFT.T  # numerical integration
+    aX2 = X.real**2 + X.imag.real**2
+    E_X = np.sum(np.sum(aX2, axis=-1) * SFT.delta_t, axis=-1) * SFT.delta_f
+    xp_assert_close(E_X, E_x, atol=atol)
+
+    # Test with random signal
+    np.random.seed(2392795)
+    x = np.random.randn(2, N_x)
+    X = SFT.stft(x)
+    xp = SFT.istft(X, k1=N_x)
+
+    assert X.shape[1] == SFT.f_pts
+    assert np.all(SFT.f >= 0.)
+    assert np.abs(max_freq - f_c) < 1.
+    xp_assert_close(x, xp, atol=atol)
+
+    # check L2-norm squared (i.e., energy) conservation:
+    E_x = np.sum(x**2, axis=-1) * SFT.T  # numeric integration
+    aX2 = X.real ** 2 + X.imag.real ** 2
+    E_X = np.sum(np.sum(aX2, axis=-1) * SFT.delta_t, axis=-1) * SFT.delta_f
+    xp_assert_close(E_X, E_x, atol=atol)
+
+    # Try with empty array
+    x = np.zeros((0, N_x))
+    X = SFT.stft(x)
+    xp = SFT.istft(X, k1=N_x)
+    assert xp.shape == x.shape