This paper describes ATR μ-Talk speech synthesis system which is designed to serve a research paradigm where system building problems can be systematically investigated by algorithmic treatments with objective measures. Non-Uniform Unit (y) selection scheme is employed to consider unit set construction and its usage simultaneously. Automatic generation, efficient construction and reduction of the unit set are discussed. The importance of Automatic Tuning of Rules for prosody control (ATR) is stressed, and some experimental results obtained through the development of the current system are presented.