diff --git a/meegkit/asr.py b/meegkit/asr.py
index 58ae53b5..05914263 100755
--- a/meegkit/asr.py
+++ b/meegkit/asr.py
@@ -288,6 +288,15 @@ def _recency_weights(counter, sample_weight):
     return weights
 
 
+def _round_half_away(x):
+    """Round half away from zero (MATLAB `round` semantics).
+
+    numpy's ``np.round`` rounds half to even (banker's rounding); this
+    matches MATLAB/Octave which round halves away from zero.
+    """
+    return np.floor(np.abs(x) + 0.5) * np.sign(x)
+
+
 def clean_windows(X, sfreq, max_bad_chans=0.2, zthresholds=[-3.5, 5],
                   win_len=.5, win_overlap=0.66, min_clean_fraction=0.25,
                   max_dropout_fraction=0.1, show=False):
@@ -354,11 +363,11 @@ def clean_windows(X, sfreq, max_bad_chans=0.2, zthresholds=[-3.5, 5],
     truncate_quant = [0.0220, 0.6000]
     step_sizes = [0.01, 0.01]
     shape_range = SHAPE_RANGE
-    max_bad_chans = np.round(X.shape[0] * max_bad_chans)
+    max_bad_chans = _round_half_away(X.shape[0] * max_bad_chans)
 
     # set data indices
     [nc, ns] = X.shape
-    N = int(win_len * sfreq)
+    N = int(_round_half_away(win_len * sfreq))
     N_raw = win_len * sfreq  # non-truncated N, avoids step-size phase drift
     offsets = np.round(np.arange(0, ns - N, N_raw * (1 - win_overlap)))
     offsets = offsets.astype(int)
@@ -543,7 +552,7 @@ def asr_calibrate(X, sfreq, cutoff=5, blocksize=100, win_len=0.5,
     X, _zf = yulewalk_filter(X, sfreq, ab=None)
 
     # window length for calculating thresholds
-    N = int(np.round(win_len * sfreq))
+    N = int(_round_half_away(win_len * sfreq))
     if ns < N:
         raise ValueError(
             f"Calibration data has {ns} samples, shorter than one analysis "
diff --git a/tests/test_asr.py b/tests/test_asr.py
index 3aa9304c..842e6586 100644
--- a/tests/test_asr.py
+++ b/tests/test_asr.py
@@ -8,7 +8,7 @@
 import pytest
 from scipy import signal
 
-from meegkit.asr import ASR, asr_calibrate, asr_process, clean_windows
+from meegkit.asr import ASR, _round_half_away, asr_calibrate, asr_process, clean_windows
 from meegkit.utils.asr import SHAPE_RANGE, fit_eeg_distribution, yulewalk, yulewalk_filter
 from meegkit.utils.covariances import block_covariance
 from meegkit.utils.matrix import sliding_window
@@ -142,6 +142,34 @@ def test_yulewalk_filter(n_chans, show=False):
         plt.show()
 
 
+def test_round_half_away():
+    """_round_half_away rounds ties away from zero, unlike numpy's banker's rounding."""
+    # exact-half (tie) behavior: rounds away from zero
+    assert _round_half_away(2.5) == 3
+    assert _round_half_away(128.5) == 129
+    assert _round_half_away(-2.5) == -3
+    assert _round_half_away(0.5) == 1
+    assert _round_half_away(1.5) == 2  # np.round gives 2 here too (even)
+
+    # divergence from banker's rounding on ties
+    assert _round_half_away(2.5) != np.round(2.5)      # 3 != 2
+    assert _round_half_away(128.5) != np.round(128.5)  # 129 != 128
+
+    # non-tie values round normally
+    assert _round_half_away(2.4) == 2
+    assert _round_half_away(2.6) == 3
+    assert _round_half_away(-2.6) == -3
+
+    # int() of the result is exact for scalars
+    assert int(_round_half_away(2.5)) == 3
+    assert int(_round_half_away(-2.5)) == -3
+
+    # array input
+    ties = np.array([2.5, 128.5, -2.5, 0.5, 1.5])
+    expected = np.array([3, 129, -3, 1, 2])
+    np.testing.assert_array_equal(_round_half_away(ties), expected)
+
+
 def test_asr_functions(show=False, method="riemann"):
     """Test ASR functions (offline use).