diff --git a/src/vocalinux/speech_recognition/recognition_manager.py b/src/vocalinux/speech_recognition/recognition_manager.py
index 3abe9c0..46b57c8 100644
--- a/src/vocalinux/speech_recognition/recognition_manager.py
+++ b/src/vocalinux/speech_recognition/recognition_manager.py
@@ -1266,7 +1266,16 @@ class SpeechRecognitionManager:
         elif has_gpu_libs:
             n_threads = max(1, multiprocessing.cpu_count() // 4)
         else:
-            n_threads = min(multiprocessing.cpu_count(), 8)
+            # Leave a couple of cores for the OS and audio capture, and avoid
+            # oversubscribing hybrid CPUs. Using *every* core drags the slow
+            # efficiency cores into the pool; whisper.cpp synchronises all threads
+            # at each decode step, so the barrier stalls on the slowest core and
+            # throughput collapses. On an 8-core hybrid CPU (Intel Core Ultra,
+            # P+E cores) the old `min(cpu_count, 8)` picked 8 = every core and
+            # measured RTF jumped from ~0.14x (6 threads) to ~2.1x (8 threads) —
+            # a ~15x slowdown. Cap at 8; base/small models rarely scale past that.
+            # Users can still override with the `whispercpp_n_threads` setting.
+            n_threads = max(1, min(multiprocessing.cpu_count() - 2, 8))
 
         load_start_time = time.time()
         loaded_backend = backend