This article introduces a granular synthesis algorithm in which grains of sound are processed in the frequency domain and combined at spectral level. In particular, each grain contributes to the sound synthesis with its magnitude spectrum only. Phase is reconstructed using spectrogram inversion techniques which support real-time computation. With this method, windowing and overlapping grains of regular size is no longer required. The algorithm presents a large number of parameters available for the dynamic manipulation of the synthetic timbre. Algorithm implementations are detailed in the paper, including comparison with traditional time-domain against the proposed frequency-domain granula-tion strategies.