Due to hardware developments in the last decade, the high-frequency end of the frequency band of seismic waves analyzed for source mechanisms has been extended into the audio-frequency range (>20 Hz). In principle, the short wavelengths corresponding to these frequencies can provide information about the details of seismic sources, but in fact, much of the "signal" is the site response of the nearsurface. Several examples of waveform data recorded at "hard rock" sites, which are generally assumed to have a "flat" transfer function, are presented to demonstrate the severe signal distortions, including fmax, produced by near-surface structures. Analysis of the geology of a number of sites indicates that the overall attenuation of high-frequency (>1 Hz) seismic waves is controlled by the whole-path-Q between source and receiver but the presence of distinct fmax site resonance peaks is controlled by the nature of the surface layer and the underlying near-surface structure. Models of vertical decoupling of the surface and nearsurface and horizontal decoupling of adjacent sites on hard rock outcrops are proposed and their behaviour is compared to the observations of hard rock site response. The upper bound to the frequency band of the seismic waves that contain significant source information which can be deconvolved from a site response or an array response is discussed in terms of fmax and the correlation of waveform distortion with the outcrop-scale geologic structure of hard rock sites. It is concluded that although the velocity structures of hard rock sites, unlike those of alluvium sites, allow some audio-frequency seismic energy to propagate to the surface, the resulting signals are a highly distorted, limited subset of the source spectra. ?? 1988 Birkha??user Verlag.