diff --git a/app/services/qwen3_websocket_asr.py b/app/services/qwen3_websocket_asr.py
index fcd0bf4..3aba177 100644
--- a/app/services/qwen3_websocket_asr.py
+++ b/app/services/qwen3_websocket_asr.py
@@ -267,6 +267,8 @@ class ConnectionContext:
     last_partial_raw_text: str = ""
     last_partial_display_text: str = ""
     stable_partial_prefix: str = ""
+    pending_partial_revision_text: str = ""
+    pending_partial_revision_rounds: int = 0
     segment_observed_text: str = ""
     segment_observed_language: str = ""
     best_partial_text: str = ""
@@ -1349,6 +1351,8 @@ class Qwen3ASRService:
         ctx.last_partial_raw_text = ""
         ctx.last_partial_display_text = ""
         ctx.stable_partial_prefix = ""
+        ctx.pending_partial_revision_text = ""
+        ctx.pending_partial_revision_rounds = 0
         ctx.segment_observed_text = ""
         ctx.segment_observed_language = ""
         ctx.best_partial_text = ""
@@ -1696,6 +1700,8 @@ class Qwen3ASRService:
         if not text:
             ctx.last_partial_raw_text = ""
             ctx.stable_partial_prefix = ""
+            ctx.pending_partial_revision_text = ""
+            ctx.pending_partial_revision_rounds = 0
             return ""
 
         previous_raw = ctx.last_partial_raw_text
@@ -1719,9 +1725,30 @@ class Qwen3ASRService:
                 1,
             )
             if len(stable_prefix) - stable_common >= tolerated_divergence and previous_emitted:
-                return previous_emitted
-            stable_prefix = text[:stable_common]
-            stable_prefix = text[: self._stable_prefix_cutoff(text, len(stable_prefix))]
+                # A single shorter or divergent snapshot may be a transient model
+                # revision. Hold the visible text until the same correction repeats.
+                if text == ctx.pending_partial_revision_text:
+                    ctx.pending_partial_revision_rounds += 1
+                else:
+                    ctx.pending_partial_revision_text = text
+                    ctx.pending_partial_revision_rounds = 1
+
+                if ctx.pending_partial_revision_rounds < 2:
+                    ctx.last_partial_raw_text = text
+                    return previous_emitted
+
+                # The repeated revision is stable enough to replace the old prefix.
+                stable_prefix = text[: self._stable_prefix_cutoff(text, stable_common)]
+                ctx.pending_partial_revision_text = ""
+                ctx.pending_partial_revision_rounds = 0
+            else:
+                ctx.pending_partial_revision_text = ""
+                ctx.pending_partial_revision_rounds = 0
+                stable_prefix = text[:stable_common]
+                stable_prefix = text[: self._stable_prefix_cutoff(text, len(stable_prefix))]
+        else:
+            ctx.pending_partial_revision_text = ""
+            ctx.pending_partial_revision_rounds = 0
 
         ctx.stable_partial_prefix = stable_prefix
         ctx.last_partial_raw_text = text
@@ -2789,6 +2816,13 @@ class Qwen3ASRService:
                             )
                             observed = update_observed_text(ctx, current, current_language)
                             visible_observed = self._trim_previous_segment_overlap(ctx, observed)
+                            if has_native_stream_snapshot:
+                                # Qwen returns full-utterance snapshots; stabilize the
+                                # displayed prefix while allowing repeated corrections.
+                                visible_observed = self._stabilize_partial_text(
+                                    ctx,
+                                    visible_observed,
+                                )
                             partial_display = self._clip_partial_text(visible_observed, ctx)
                             if (
                                 partial_display
@@ -2934,6 +2968,10 @@ class Qwen3ASRService:
                     emit_segment_start=False,
                 )
 
+            # Drain pending diarization before the final event; the browser closes
+            # the socket on `end`, so later speaker updates would otherwise be lost.
+            await ctx.speaker_job_queue.join()
+
             final_updates = self._recluster_confirmed_segments(ctx)
             for updated in final_updates:
                 sentence_payload = self._build_tencent_sentence(