Mary sounds quite good. Apparently if you know how to write input files in the correct format you can get it to speak much more like a real human by including stress/speed/tone/etc. variations too.