Twenty years of work with sinusoidal modeling of speech has lead to very competitive principles of low rate coding. In this study, we discuss a few issues in the design of a sinusoidal coding system. We stress that by a careful design of all blocks of the encoder and decoder, allowing for some additional complexity, it is possible to build a low rate coder free of many of the artifacts associated with conventional vocoding systems. We focus this discussion on multi-band partial voicing and on parameter smoothing and interpolation.