diff --git a/ocrolib/lstm.py b/ocrolib/lstm.py
index bf10ab56..42793d49 100644
--- a/ocrolib/lstm.py
+++ b/ocrolib/lstm.py
@@ -852,12 +852,15 @@ def ctc_align_targets(outputs,targets,threshold=100.0,verbose=0,debug=0,lo=1e-5)
 def normalize_nfkc(s):
     return unicodedata.normalize('NFKC',s)
 
+def normalize_nfc(s):
+    return unicodedata.normalize('NFC',s)
+
 def add_training_info(network):
     return network
 
 class SeqRecognizer:
     """Perform sequence recognition using BIDILSTM and alignment."""
-    def __init__(self,ninput,nstates,noutput=-1,codec=None,normalize=normalize_nfkc):
+    def __init__(self,ninput,nstates,noutput=-1,codec=None,normalize=normalize_nfc):
         self.Ni = ninput
         if codec: noutput = codec.size()
         assert noutput>0
diff --git a/ocropus-rtrain b/ocropus-rtrain
index 9d365ae3..bec093b2 100755
--- a/ocropus-rtrain
+++ b/ocropus-rtrain
@@ -134,7 +134,7 @@ if args.codec!=[]:
     print(args.codec)
     for fname in ocrolib.glob_all(args.codec):
         transcript = ocrolib.read_text(fname)
-        l = list(lstm.normalize_nfkc(transcript))
+        l = list(lstm.normalize_nfc(transcript))
         charset = charset.union(l)
     charset = sorted(list(charset))
     charset = [c for c in charset if c>" " and c!="~"]
@@ -172,7 +172,7 @@ def load_lstm(fname):
     if args.clstm:
         network = lstm.SeqRecognizer(args.height,args.hiddensize,
             codec=codec,
-            normalize=lstm.normalize_nfkc)
+            normalize=lstm.normalize_nfc)
         import clstm
         mylstm = clstm.make_BIDILSTM()
         mylstm.init(network.No,args.hiddensize,network.Ni)
@@ -193,7 +193,7 @@ else:
     last_save = None
     network = lstm.SeqRecognizer(args.height,args.hiddensize,
         codec=codec,
-        normalize=lstm.normalize_nfkc)
+        normalize=lstm.normalize_nfc)
     if args.clstm:
         import clstm
         mylstm = clstm.make_BIDILSTM()