diff --git a/CHANGELOG.md b/CHANGELOG.md
index d33d477ef8d..4bc97e58972 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -39,6 +39,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Fixed `Metric` ignoring an active `torch.device` context manager on torch 2.3-2.7 ([#3448](https://github.com/Lightning-AI/torchmetrics/pull/3448))
 
 
+- Fixed swapped hypothesis and reference arguments in `TER`, which made an empty hypothesis score a perfect 0.0 ([#3479](https://github.com/Lightning-AI/torchmetrics/pull/3479))
+
+
 ---
 
 ## [1.9.0] - 2026-03-05
diff --git a/src/torchmetrics/functional/text/ter.py b/src/torchmetrics/functional/text/ter.py
index 2d7a6211e0d..40e32fb96cf 100644
--- a/src/torchmetrics/functional/text/ter.py
+++ b/src/torchmetrics/functional/text/ter.py
@@ -444,7 +444,7 @@ def _compute_sentence_statistics(pred_words: list[str], target_words: list[list[
     best_num_edits = tensor(2e16)
 
     for tgt_words in target_words:
-        num_edits = _translation_edit_rate(tgt_words, pred_words)
+        num_edits = _translation_edit_rate(pred_words, tgt_words)
         tgt_lengths += len(tgt_words)
         if num_edits < best_num_edits:
             best_num_edits = num_edits
diff --git a/tests/unittests/text/test_ter.py b/tests/unittests/text/test_ter.py
index 5b51d8da704..1a46ca97e29 100644
--- a/tests/unittests/text/test_ter.py
+++ b/tests/unittests/text/test_ter.py
@@ -176,3 +176,22 @@ def test_ter_return_sentence_level_class():
     targets = _inputs_single_sentence_multiple_references.target
     _, sentence_ter = ter_metric(preds, targets)
     isinstance(sentence_ter, Tensor)
+
+
+def test_ter_empty_hypothesis_is_not_a_perfect_score():
+    """Test that a hypothesis producing nothing is charged one deletion per reference word."""
+    assert translation_edit_rate([""], [["hello world foo"]]) == tensor(1.0)
+
+
+def test_ter_matches_reference_when_shifts_are_asymmetric():
+    """Test that edits are counted from hypothesis to reference, not the other way around.
+
+    The shift search is not symmetric, so swapping the two sides changes the number of edits found.
+
+    """
+    preds = ["hello hello the a dog"]
+    targets = [["jumps dog lazy the"]]
+    expected = _reference_sacrebleu_ter(
+        preds, targets, normalized=False, no_punct=False, asian_support=False, case_sensitive=True
+    )
+    assert translation_edit_rate(preds, targets, lowercase=False) == expected