diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py
index fe0d0a872..df1e52202 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py
@@ -101,6 +101,13 @@ class AsrDataModule:
             "collect the batches.",
         )
 
+        group.add_argument(
+            "--on-the-fly-num-workers",
+            type=int,
+            default=0,
+            help="The number of workers for on-the-fly feature extraction",
+        )
+
         group.add_argument(
             "--enable-spec-aug",
             type=str2bool,
@@ -212,7 +219,10 @@ class AsrDataModule:
         train = K2SpeechRecognitionDataset(
             cut_transforms=transforms,
             input_strategy=(
-                OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+                OnTheFlyFeatures(
+                    extractor=Fbank(FbankConfig(num_mel_bins=80)),
+                    num_workers=self.args.on_the_fly_num_workers,
+                )
                 if on_the_fly_feats
                 else PrecomputedFeatures()
             ),