diff --git a/training/deepspeech_training/util/signal_augmentations.py b/training/deepspeech_training/util/signal_augmentations.py
index f9156c39ac..d69622d435 100644
--- a/training/deepspeech_training/util/signal_augmentations.py
+++ b/training/deepspeech_training/util/signal_augmentations.py
@@ -141,7 +141,8 @@ def apply(self, sample, clock):
         n_gaps = pick_value_from_range(self.n_gaps, clock=clock)
         for _ in range(n_gaps):
             size = pick_value_from_range(self.size, clock=clock)
-            offset = max(0, random.randint(0, len(audio) - size - 1))
+            size = min(size, len(audio) // 10)  # a gap should never exceed 10 percent of the audio
+            offset = random.randint(0, max(0, len(audio) - size - 1))
             audio[offset:offset + size] = 0
         sample.audio = audio