diff --git a/modules/ltx/ltx_process.py b/modules/ltx/ltx_process.py
index 2fc79922c..5cbe5cb0c 100644
--- a/modules/ltx/ltx_process.py
+++ b/modules/ltx/ltx_process.py
@@ -360,8 +360,11 @@ def run_ltx(task_id,
                         yield from abort('Video: process_images returned no frames', ok=True, p=p)
                         return
                     pixels = processed.images
-                    if getattr(processed, 'audio', None) is not None:
-                        audio = processed.audio
+                    raw_audio = getattr(processed, 'audio', None)
+                    if raw_audio is not None:
+                        # Strip batch dim from (B, 2, N); write_audio expects (2, N) for the
+                        # transpose-to-interleaved path used by AAC s16.
+                        audio = raw_audio[0].float().cpu() if raw_audio.ndim == 3 else raw_audio.float().cpu()
                     latents = None
             except AssertionError as e:
                 yield from abort(e, ok=True, p=p)
diff --git a/modules/processing.py b/modules/processing.py
index 0306b20a8..36c600ebe 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -528,7 +528,7 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
                         output_images.append(batch_image)
                         infotexts.append(batch_infotext)
 
-            audio = getattr(samples, 'audio', None)
+            audio = getattr(samples, 'audio', None) or getattr(p, 'audio_capture', None)
 
             if shared.cmd_opts.lowvram:
                 devices.torch_gc(force=True, reason='lowvram')
diff --git a/modules/processing_diffusers.py b/modules/processing_diffusers.py
index b7b1030ef..6b701a966 100644
--- a/modules/processing_diffusers.py
+++ b/modules/processing_diffusers.py
@@ -628,6 +628,10 @@ def process_diffusers(p: processing.StableDiffusionProcessing):
     timer.process.add('lora', lora_common.timer.total)
     lora_common.timer.clear(complete=True)
 
+    # process_decode flattens video output to a frame list and drops the audio attribute;
+    # stash it on `p` so video pipelines can recover it after process_images returns.
+    if output is not None and getattr(output, 'audio', None) is not None:
+        p.audio_capture = output.audio
     results = process_decode(p, output)
     timer.process.record('decode')