In this work, a novel subspace-based algorithm is presented for automated random noise reduction in online recorded music. Musical signal enhancement is a separate issue from the well-studied speech enhancement problem due to the particularly wide range of signal characteristics encountered, and thus requires a very general approach. Because similar issues drive denoising advances in seismic signal processing, it is argued that an algorithm can be developed through a cross-disciplinary approach. Inspired by an enhancement method for seismic sections, noise reduction is achieved by applying a singular value decomposition-based image enhancement technique, known as eigenimage filtering, to the time-frequency representation of the musical signal. Classic eigenimage filtering approximates a full-rank matrix by its closest rank-deficient approximation; the preserved and discarded parts of the matrix correspond to the signal and noise subspaces, respectively. Under the assumption of a quasi-stationary signal, this technique is applied to the short-time Fourier transform of the signal. However, because the standard eigenimage filtering approach results in unwanted residual noise characteristics when applied in this domain, an adapted version of the technique is used. In this adaptation, all singular values are altered but none are set to zero, and the alteration is dependent on the singular values encountered. Therefore, the method is data-adaptive. Subjective and objective performance measures indicate that the method is capable of improving the quality of noisy recordings, and that its quality is competitive compared with an open-source noise reduction algorithm whilst having the advantages of automation and fewer user-defined parameters.