A signal processing device includes a prior distribution storage that stores a prior distribution group that is a set of prior distributions representing representative spectral shapes of spectral bases of an acoustic event specified as a detection target, and a basis generation unit that, using as input a spectrogram for basis generation, generates an acoustic event basis group that is a set of the spectral bases of the acoustic event specified as a detection target. The basis generation unit performs sound source separation on the spectrogram for basis generation and thereby generates an acoustic event basis group, using respective prior distributions included in a prior distribution group as a constraint for spectral bases of a corresponding acoustic event.