Sample-based synthesis is one of the most common synthesis methods of the digital synthesizer. It can synthesize musical instrument sounds well because it plays recorded sounds instead of generating sounds such as additive and frequency modulation syntheses. The recorded sounds are usually stored in the time domain. If the recorded sounds are stored in the time-frequency domain, sound processing can be simplified. Users can process audio signals intuitively by operating the time-frequency bins of the spectrograms. However, the representation of sounds in the time-frequency domain usually requires more memory than that in the time domain. In synthesizers, if much data is needed, hardware costs are increased correspondingly. Therefore, a technique to reduce the amount of data is required. In this paper, to reduce the amount of data, we introduce a sparse modeling technique for musical instrument sounds in the time-frequency domain. We propose an optimization algorithm of sparse modeling with four shrinkage operators. Numerical experiments show that the data quantity of the signals reduced by over 95 percent by using the technique.