The timbre of an instrument is usually represented by sinusoids plus noise. Spectral modeling synthesis (SMS) is an audio synthesis technique which can create musical timbre and give control over the frequency and amplitude. Additive synthesis and LPC synthesis are usually applied for synthesizing sinusoids and residuals, respectively. However, it takes fairly large computing power while implementing the algorithms. The purpose of this paper is to present GPU-based techniques of implementing SMS for real-time audio processing by using parallelism and programmability in graphics pipeline. The performance is compared to CPU-based implementations.