diff --git a/bots/paused_caption_filter.py b/bots/paused_caption_filter.py
new file mode 100644
index 000000000..c43e7d9ba
--- /dev/null
+++ b/bots/paused_caption_filter.py
@@ -0,0 +1,23 @@
+class PausedCaptionFilter:
+    """Drops caption updates that would leak speech from a recording pause.
+
+    Platform captions (Google Meet/Teams/Zoom web) arrive as *growing* segment texts keyed
+    by ``deviceId:captionId``. Ignoring updates only while paused is not enough:
+    the first update after resume carries the whole segment, including the words
+    spoken during the pause. Any segment that sends an update during a pause is
+    therefore frozen for good; segments that start after resume pass normally.
+    """
+
+    def __init__(self):
+        self._frozen_keys = set()
+
+    @staticmethod
+    def key(caption):
+        return f"{caption.get('deviceId')}:{caption.get('captionId')}"
+
+    def should_drop(self, caption, recording_paused):
+        key = self.key(caption)
+        if recording_paused:
+            self._frozen_keys.add(key)
+            return True
+        return key in self._frozen_keys
diff --git a/bots/tests/test_paused_caption_filter.py b/bots/tests/test_paused_caption_filter.py
new file mode 100644
index 000000000..f8805f5cc
--- /dev/null
+++ b/bots/tests/test_paused_caption_filter.py
@@ -0,0 +1,31 @@
+import unittest
+
+from bots.paused_caption_filter import PausedCaptionFilter
+
+
+class PausedCaptionFilterTest(unittest.TestCase):
+    def test_segment_spanning_a_pause_stays_dropped_after_resume(self):
+        f = PausedCaptionFilter()
+        seg = {"captionId": "c1", "deviceId": "u1"}
+        self.assertFalse(f.should_drop(seg, recording_paused=False))  # before pause: recorded
+        self.assertTrue(f.should_drop(seg, recording_paused=True))  # during pause: dropped
+        # After resume the same segment carries the paused words -> still dropped
+        self.assertTrue(f.should_drop(seg, recording_paused=False))
+
+    def test_new_segment_after_resume_passes(self):
+        f = PausedCaptionFilter()
+        f.should_drop({"captionId": "c1", "deviceId": "u1"}, recording_paused=True)
+        self.assertFalse(f.should_drop({"captionId": "c2", "deviceId": "u1"}, recording_paused=False))
+
+    def test_segments_are_keyed_per_device(self):
+        f = PausedCaptionFilter()
+        f.should_drop({"captionId": "c1", "deviceId": "u1"}, recording_paused=True)
+        self.assertFalse(f.should_drop({"captionId": "c1", "deviceId": "u2"}, recording_paused=False))
+
+    def test_untouched_segments_pass_when_not_paused(self):
+        f = PausedCaptionFilter()
+        self.assertFalse(f.should_drop({"captionId": "c1", "deviceId": "u1"}, recording_paused=False))
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/bots/web_bot_adapter/web_bot_adapter.py b/bots/web_bot_adapter/web_bot_adapter.py
index 16ba2c0b0..2ab7614c6 100644
--- a/bots/web_bot_adapter/web_bot_adapter.py
+++ b/bots/web_bot_adapter/web_bot_adapter.py
@@ -25,6 +25,7 @@
 from bots.automatic_leave_utils import participant_is_another_bot
 from bots.bot_adapter import BotAdapter
 from bots.models import ParticipantEventTypes, RecordingViews
+from bots.paused_caption_filter import PausedCaptionFilter
 from bots.per_participant_realtime_video_configuration import PerParticipantRealtimeVideoConfiguration
 from bots.room_sync_source_participant_configuration import RoomSyncSourceParticipantConfiguration
 from bots.room_sync_utils import add_bot_indicator_to_display_name
@@ -132,6 +133,7 @@ def __init__(
         self.ready_to_send_chat_messages = False
 
         self.recording_paused = False
+        self.paused_caption_filter = PausedCaptionFilter()
 
     def pause_recording(self):
         self.recording_paused = True
@@ -378,7 +380,8 @@ def handle_failed_to_join(self, reason):
         self.subclass_specific_handle_failed_to_join(reason)
 
     def handle_caption_update(self, json_data):
-        if self.recording_paused:
+        # Segments that carry speech from a pause stay dropped after resume, too.
+        if self.paused_caption_filter.should_drop(json_data["caption"], self.recording_paused):
             return
 
         # Count a caption as audio activity
diff --git a/version.json b/version.json
index 21dff1d57..8678376ce 100644
--- a/version.json
+++ b/version.json
@@ -1,3 +1,3 @@
 {
-  "version": "1.82.6"
+  "version": "1.82.7"
 }