diff --git a/sms_wsj/database/wsj/create_json.py b/sms_wsj/database/wsj/create_json.py
index 4b89a47..7912526 100644
--- a/sms_wsj/database/wsj/create_json.py
+++ b/sms_wsj/database/wsj/create_json.py
@@ -8,6 +8,7 @@
 import os
 import re
 import tempfile
+import subprocess
 from pathlib import Path
 
 import sacred
@@ -182,18 +183,20 @@ def normalize_transcription(transcriptions, wsj_root: Path):
     :param wsj_root: Path to WSJ database
 
     :return result: Clean transcription dictionary
+
+    >>> transcriptions = {'ID1': 'Hello World, and bye!?', 'ID2': 'What? Yes.'}
+    >>> normalize_transcription(transcriptions, '')
+    {'ID1': 'HELLO WORLD, AND BYE!?', 'ID2': 'WHAT? YES.'}
+
     """
     assert len(transcriptions) > 0, 'No transcriptions to clean up.'
-    with tempfile.TemporaryDirectory() as temporary_directory:
-        temporary_directory = Path(temporary_directory).absolute()
-        with open(temporary_directory / 'dirty.txt', 'w') as f:
-            for key, value in transcriptions.items():
-                f.write('{} {}\n'.format(key, value))
-        result = sh.perl(
-            sh.cat(str(temporary_directory / 'dirty.txt')),
-            kaldi_wsj_tools / 'normalize_transcript.pl',
-            '<NOISE>'
-        )
+
+    text = ''.join([f'{key} {value}\n' for key, value in transcriptions.items()])
+    cp = subprocess.run(
+        ['perl', kaldi_wsj_tools / 'normalize_transcript.pl', '<NOISE>'],
+        input=text, stdout=subprocess.PIPE, check=True, universal_newlines=True)
+    result = cp.stdout
+
     result = [line.split(maxsplit=1) for line in result.strip().split('\n')]
     result = {k: v for k, v in result}
     return result