diff --git a/lib/matplotlib/mlab.py b/lib/matplotlib/mlab.py
index 3552904c3d74..195f8cf01393 100644
--- a/lib/matplotlib/mlab.py
+++ b/lib/matplotlib/mlab.py
@@ -426,8 +426,9 @@ def _spectral_helper(x, y=None, NFFT=None, Fs=None, detrend_func=None,
             # windowing loss; see Bendat & Piersol Sec 11.5.2.
             result /= (np.abs(window)**2).sum()
         else:
-            # In this case, preserve power in the segment, not amplitude
-            result /= np.abs(window).sum()**2
+            # Preserve segment power using the signed window sum, matching
+            # SciPy's welch(scale='spectrum') normalization.
+            result /= (window.sum())**2
 
     t = np.arange(NFFT/2, len(x) - NFFT/2 + 1, NFFT - noverlap)/Fs
 
diff --git a/lib/matplotlib/tests/test_mlab_window_correction.py b/lib/matplotlib/tests/test_mlab_window_correction.py
new file mode 100644
index 000000000000..2604e594ae35
--- /dev/null
+++ b/lib/matplotlib/tests/test_mlab_window_correction.py
@@ -0,0 +1,113 @@
+import numpy as np
+import pytest
+from numpy.testing import assert_allclose
+
+from matplotlib import mlab
+
+
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:'enablePackrat' deprecated - use 'enable_packrat':DeprecationWarning"
+)
+
+
+_FS = 1024.0
+_NFFT = 512
+_F0 = 128.0
+
+
+def _test_signal():
+    t = np.arange(_NFFT) / _FS
+    return np.sin(2 * np.pi * _F0 * t)
+
+
+def _negative_value_window_values(length):
+    n = np.arange(length, dtype=float)
+    return 0.5 + np.cos(2 * np.pi * n / (length - 1))
+
+
+def _negative_value_window(segment):
+    weights = _negative_value_window_values(segment.shape[0])
+    return weights * segment
+
+
+def _expected_psd(window_vals, scale_by_freq):
+    segments = mlab._stride_windows(_test_signal(), _NFFT, noverlap=0)
+    segments = mlab.detrend(segments, mlab.detrend_none, axis=0)
+    windowed = segments * window_vals[:, None]
+
+    fft = np.fft.fft(windowed, n=_NFFT, axis=0)
+    num_freqs = _NFFT // 2 + 1
+    fft = fft[:num_freqs, :]
+
+    periodogram = np.conj(fft) * fft
+    periodogram[1:-1] *= 2.0
+
+    if scale_by_freq:
+        periodogram /= _FS
+        periodogram /= (np.abs(window_vals) ** 2).sum()
+    else:
+        periodogram /= (window_vals.sum()) ** 2
+
+    return periodogram.mean(axis=1)
+
+
+def test_psd_spectrum_uses_window_sum_square():
+    window_vals = _negative_value_window_values(_NFFT)
+    spectrum, freqs = mlab.psd(
+        x=_test_signal(),
+        NFFT=_NFFT,
+        Fs=_FS,
+        window=_negative_value_window,
+        noverlap=0,
+        pad_to=_NFFT,
+        sides="onesided",
+        scale_by_freq=False,
+    )
+
+    expected = _expected_psd(window_vals, scale_by_freq=False)
+    assert_allclose(spectrum, expected, rtol=1e-12, atol=1e-12)
+
+
+def test_psd_density_uses_window_energy_norm():
+    window_vals = _negative_value_window_values(_NFFT)
+    density, _ = mlab.psd(
+        x=_test_signal(),
+        NFFT=_NFFT,
+        Fs=_FS,
+        window=_negative_value_window,
+        noverlap=0,
+        pad_to=_NFFT,
+        sides="onesided",
+        scale_by_freq=True,
+    )
+
+    expected = _expected_psd(window_vals, scale_by_freq=True)
+    assert_allclose(density, expected, rtol=1e-12, atol=1e-12)
+
+
+def test_psd_density_spectrum_consistency():
+    window_vals = _negative_value_window_values(_NFFT)
+    density, freqs = mlab.psd(
+        x=_test_signal(),
+        NFFT=_NFFT,
+        Fs=_FS,
+        window=_negative_value_window,
+        noverlap=0,
+        pad_to=_NFFT,
+        sides="onesided",
+        scale_by_freq=True,
+    )
+    spectrum, _ = mlab.psd(
+        x=_test_signal(),
+        NFFT=_NFFT,
+        Fs=_FS,
+        window=_negative_value_window,
+        noverlap=0,
+        pad_to=_NFFT,
+        sides="onesided",
+        scale_by_freq=False,
+    )
+
+    lhs = density * (window_vals**2).sum()
+    rhs = spectrum / _FS * (window_vals.sum()) ** 2
+    assert_allclose(lhs, rhs, rtol=1e-12, atol=1e-12)