Recently we (Trac XmlRpcPlugin maintainers) have been notified of the fact that the regular expression above strips surrogate pairs on Python narrow builds (see th:comment:13:ticket:11050) . An alternative approach consists in using the following regex (see th:changeset:13729) .

_illegal_unichrs = [(0x00, 0x08), (0x0B, 0x0C), (0x0E, 0x1F), 
                        (0x7F, 0x84), (0x86, 0x9F), 
                        (0xFDD0, 0xFDDF), (0xFFFE, 0xFFFF)] 
if sys.maxunicode >= 0x10000:  # not narrow build 
        _illegal_unichrs.extend([(0x1FFFE, 0x1FFFF), (0x2FFFE, 0x2FFFF), 
                                 (0x3FFFE, 0x3FFFF), (0x4FFFE, 0x4FFFF), 
                                 (0x5FFFE, 0x5FFFF), (0x6FFFE, 0x6FFFF), 
                                 (0x7FFFE, 0x7FFFF), (0x8FFFE, 0x8FFFF), 
                                 (0x9FFFE, 0x9FFFF), (0xAFFFE, 0xAFFFF), 
                                 (0xBFFFE, 0xBFFFF), (0xCFFFE, 0xCFFFF), 
                                 (0xDFFFE, 0xDFFFF), (0xEFFFE, 0xEFFFF), 
                                 (0xFFFFE, 0xFFFFF), (0x10FFFE, 0x10FFFF)]) 

_illegal_ranges = ["%s-%s" % (unichr(low), unichr(high)) 
                   for (low, high) in _illegal_unichrs] 
_illegal_xml_chars_RE = re.compile(u [%s]  % u  .join(_illegal_ranges)) 

p.s. See this post on surrogates explaining what they are for .

Update so as to not to match (replace) 0x0D which is a valid XML character.

Here s an updated version of Olemis Lang s answer for Python 3:

import re
import sys

illegal_unichrs = [(0x00, 0x08), (0x0B, 0x0C), (0x0E, 0x1F),
                   (0x7F, 0x84), (0x86, 0x9F),
                   (0xFDD0, 0xFDDF), (0xFFFE, 0xFFFF)]
if sys.maxunicode >= 0x10000:  # not narrow build
    illegal_unichrs.extend([(0x1FFFE, 0x1FFFF), (0x2FFFE, 0x2FFFF),
                            (0x3FFFE, 0x3FFFF), (0x4FFFE, 0x4FFFF),
                            (0x5FFFE, 0x5FFFF), (0x6FFFE, 0x6FFFF),
                            (0x7FFFE, 0x7FFFF), (0x8FFFE, 0x8FFFF),
                            (0x9FFFE, 0x9FFFF), (0xAFFFE, 0xAFFFF),
                            (0xBFFFE, 0xBFFFF), (0xCFFFE, 0xCFFFF),
                            (0xDFFFE, 0xDFFFF), (0xEFFFE, 0xEFFFF),
                            (0xFFFFE, 0xFFFFF), (0x10FFFE, 0x10FFFF)])

illegal_ranges = [fr {chr(low)}-{chr(high)}  for (low, high) in illegal_unichrs]
xml_illegal_character_regex =  [  +   .join(illegal_ranges) +  ] 
illegal_xml_chars_re = re.compile(xml_illegal_character_regex)
# filtered_string = illegal_xml_chars_re.sub(  , original_string)

You could also use unicode s translate method to delete selected codepoints. However, the mapping you have is pretty big (2128 codepoints) and that might make it much slower than just using a regex:

ranges = [(0, 8), (0xb, 0x1f), (0x7f, 0x84), (0x86, 0x9f), (0xd800, 0xdfff), (0xfdd0, 0xfddf), (0xfffe, 0xffff)]
# fromkeys creates  the wanted (codepoint -> None) mapping
nukemap = dict.fromkeys(r for start, end in ranges for r in range(start, end+1))
clean = dirty.translate(nukemap)