We are having same problem, but in our case we don't use localized templates
but have utf-8 characters inside MySQL which get double utf-8 encoded before
they are sent to browser.
If you feel brave, you can try attached patch with fixes problem for us.
Below is short summary of changes from patch:
This patch tries to clean up utf-8 handling in Koha.
In current implementation (mostly commented out in this patch)
uses heuristic to guess which strings need decoding from utf-8
to binary representation and doesn't support utf-8 characters
in templates and has problems with utf-8 data from database.
With this changes, Koha perl code always uses utf-8 encoding
correctly. All incomming data from database is allready
correctly marked as utf-8, and decoding of utf8 is required
only from Zebra and XSLT transfers which don't set utf-8 flag
For output, standard perl :utf8 handler is used removing various
"wide character" warnings as side-effect.