diff --git a/tests/test_subtitle_pauses.py b/tests/test_subtitle_pauses.py
new file mode 100644
index 000000000..155268fa4
--- /dev/null
+++ b/tests/test_subtitle_pauses.py
@@ -0,0 +1,23 @@
+import pytest
+
+from whisper.utils import WriteVTT
+
+
+@pytest.mark.parametrize(
+    "next_start, expected_text",
+    [
+        (3.5, ["hello again"]),  # 2.5 seconds of silence
+        (4.0, ["hello again"]),  # exactly the three-second threshold
+        (4.1, ["hello", "again"]),  # more than three seconds of silence
+    ],
+)
+def test_subtitle_break_uses_gap_between_words(next_start, expected_text):
+    words = [
+        {"start": 0.0, "end": 1.0, "word": " hello"},
+        {"start": next_start, "end": next_start + 0.5, "word": " again"},
+    ]
+    result = {"segments": [{"start": 0.0, "end": next_start + 0.5, "words": words}]}
+
+    cues = WriteVTT(".").iterate_result(result, max_line_width=40, max_line_count=2)
+
+    assert [text for _, _, text in cues] == expected_text
diff --git a/whisper/utils.py b/whisper/utils.py
index 79658fc75..bb64a9ca5 100644
--- a/whisper/utils.py
+++ b/whisper/utils.py
@@ -144,7 +144,7 @@ def iterate_subtitles():
             line_count = 1
             # the next subtitle to yield (a list of word timings with whitespace)
             subtitle: List[dict] = []
-            last: float = get_start(result["segments"]) or 0.0
+            previous_end: Optional[float] = None
             for segment in result["segments"]:
                 chunk_index = 0
                 words_count = max_words_per_line
@@ -157,7 +157,9 @@ def iterate_subtitles():
                     ):
                         timing = original_timing.copy()
                         long_pause = (
-                            not preserve_segments and timing["start"] - last > 3.0
+                            not preserve_segments
+                            and previous_end is not None
+                            and timing["start"] - previous_end > 3.0
                         )
                         has_room = line_len + len(timing["word"]) <= max_line_width
                         seg_break = i == 0 and len(subtitle) > 0 and preserve_segments
@@ -188,7 +190,7 @@ def iterate_subtitles():
                                 timing["word"] = "\n" + timing["word"]
                             line_len = len(timing["word"].strip())
                         subtitle.append(timing)
-                        last = timing["start"]
+                        previous_end = timing["end"]
                     chunk_index += max_words_per_line
             if len(subtitle) > 0:
                 yield subtitle