In this paper we present a technique for lossless compression of the sinusoidal modeling parameters. Compression is indeed useful for embedding spectral sounds in a synthesizer, broadcasting spectral sounds or simply storing many of them on a medium. This technique consists in compressing the frequency and amplitude parameters of each partial by adaptively sub-sampling and encoding their evolutions with time. It can be easily extended to handle spectral envelopes instead of partials, that is two-dimensional structures instead of one-dimensional ones. We have also designed a very compact file format for (compressed) spectral sounds based on this compression scheme.