diff --git a/Lib/subprocess.py b/Lib/subprocess.py
--- a/Lib/subprocess.py
+++ b/Lib/subprocess.py
@@ -828,8 +828,8 @@
 
 
     def _translate_newlines(self, data, encoding):
-        data = data.replace(b"\r\n", b"\n").replace(b"\r", b"\n")
-        return data.decode(encoding)
+        data = data.decode(encoding)
+        return data.replace("\r\n", "\n").replace("\r", "\n")
 
     def __enter__(self):
         return self
diff --git a/Lib/test/test_subprocess.py b/Lib/test/test_subprocess.py
--- a/Lib/test/test_subprocess.py
+++ b/Lib/test/test_subprocess.py
@@ -4,6 +4,7 @@
 import sys
 import signal
 import io
+import locale
 import os
 import errno
 import tempfile
@@ -635,6 +636,28 @@
         (stdout, stderr) = p.communicate("line1\nline3\n")
         self.assertEqual(p.returncode, 0)
 
+    def test_universal_newlines_communicate_encoding_utf16(self):
+        # Check that universal newlines mode works when the preferred
+        # encoding is UTF-16.  See issue #15595.
+        old_getpreferredencoding = locale.getpreferredencoding
+        code = ("import sys; "
+                r"sys.stdout.buffer.write('a\r\nb'.encode('utf-16'))")
+        args = [sys.executable, '-c', code]
+        try:
+            # Popen() defaults to locale.getpreferredencoding(False).
+            locale.getpreferredencoding = lambda do_setlocale: 'utf-16'
+            # We set stdin to be non-None because, as of this writing,
+            # a different code path is used when the number of pipes is
+            # zero or one.
+            popen = subprocess.Popen(args, universal_newlines=True,
+                                     stdin=subprocess.PIPE,
+                                     stdout=subprocess.PIPE)
+            stdout, stderr = popen.communicate(input='')
+        finally:
+            locale.getpreferredencoding = old_getpreferredencoding
+
+        self.assertEquals(stdout, 'a\nb')
+
     def test_no_leaking(self):
         # Make sure we leak no resources
         if not mswindows: