diff --git a/src/audio/pipewire/PwJamesDspPlugin.cpp b/src/audio/pipewire/PwJamesDspPlugin.cpp
index 42c7d4cc..4f9737b9 100644
--- a/src/audio/pipewire/PwJamesDspPlugin.cpp
+++ b/src/audio/pipewire/PwJamesDspPlugin.cpp
@@ -2,6 +2,10 @@
 #include <QString>
 #include <clocale>
 
+#if defined(__x86_64__) || defined(__i386__)
+#include <xmmintrin.h>
+#endif
+
 extern "C" {
 #ifdef DEBUG_FPE
 #include <fenv.h>
@@ -78,6 +82,20 @@ void PwJamesDspPlugin::process(float* left_in,
                                float* right_out,
                                size_t length)
 {
+#if defined(__x86_64__) || defined(__i386__)
+  // Denormals are computed ~10x slower on x86. The DSP state decays into the
+  // denormal range once the input turns into digital silence (e.g. right after
+  // a stream stops), the filter no longer fits into the graph quantum and
+  // every cycle ends in an xrun, which keeps one CPU core busy forever.
+  // Enable flush-to-zero / denormals-are-zero once per processing thread.
+  static thread_local bool denormals_flushed = false;
+
+  if (!denormals_flushed) {
+    _mm_setcsr(_mm_getcsr() | 0x8040u);  // FTZ | DAZ
+    denormals_flushed = true;
+  }
+#endif
+
   if (bypass)
   {
       memcpy(left_out, left_in, length * sizeof (float));
diff --git a/src/audio/pulseaudio/wrapper/gstjamesdsp.c b/src/audio/pulseaudio/wrapper/gstjamesdsp.c
index e3b86a8d..7da62f5d 100644
--- a/src/audio/pulseaudio/wrapper/gstjamesdsp.c
+++ b/src/audio/pulseaudio/wrapper/gstjamesdsp.c
@@ -7,6 +7,10 @@
 #include <gst/controller/controller.h>
 #include <math.h>
 
+#if defined(__x86_64__) || defined(__i386__)
+#include <xmmintrin.h>
+#endif
+
 GST_DEBUG_CATEGORY_STATIC (gst_jamesdsp_debug);
 #define GST_CAT_DEFAULT gst_jamesdsp_debug
 /* Filter signals and args */
@@ -241,6 +245,18 @@ gst_jamesdsp_transform_ip(GstBaseTransform *base, GstBuffer *buf) {
     GstClockTime timestamp, stream_time;
     GstMapInfo map;
 
+#if defined(__x86_64__) || defined(__i386__)
+    /* See PwJamesDspPlugin::process(): denormal DSP state after the input
+     * turns to digital silence is ~10x slower on x86 and makes the pipeline
+     * overrun. Enable flush-to-zero / denormals-are-zero once per thread. */
+    static __thread gboolean denormals_flushed = FALSE;
+
+    if (!denormals_flushed) {
+        _mm_setcsr(_mm_getcsr() | 0x8040u); /* FTZ | DAZ */
+        denormals_flushed = TRUE;
+    }
+#endif
+
     timestamp = GST_BUFFER_TIMESTAMP (buf);
     stream_time =
             gst_segment_to_stream_time(&base->segment, GST_FORMAT_TIME, timestamp);