The Festival Speech Synthesis System, as the name goes, is probably one of the best software-based text-to-speech speech synthesis systems that is available as open source (under a BSD-like license). It was developed at The Centre for Speech Technology Research at University of Edinburgh. It is portable and works in *NIX and Windows (sort of).
The quality of voice with 16-bit output is tolerable enough (not as good as of some commercial programs though, but tolerable!), and it supports American and British pronouncation of English, as well as other languages (Spanish and Welsh, probably more in future if people are motivated enough - people had made some severe hacks to make it speak Finnish through the Spanish samplekit =). Festival can read plain text, and also supports SABLE markup, and can be extended to support other text markups as well. It has various APIs to control it, and also has a Scheme-based command interpreter.
Personally, I've used Festival to do all sorts of odd things, to build the "Evil Computer from the Year 2000" from old bad Sci-Fi movies (self-fulfilling prophecies, indeed! And I'm watching "sports from year 2000" too!) - now, if only IBM would open-source their ViaVoice package and I could also command it... I use my XMMS InfoPipe hack to make XMMS say the title of the song when I'm in bed, the monitor is turned off, and I use the remote control to control XMMS. I also recently hacked together EveryLecture, a program that lectures on various subjects, taking its information from Everything2.
http://www.cstr.ed.ac.uk/projects/festival/
http://festvox.org/festival/