Restoration and Enhancement of Solo Guitar Recordings Based on Sound Source Modeling*

Audio enhancement is a wide concept and is closely related to audio restoration. An intuitive idea of audio enhancement is associated with any audio processing that is able to improve the perceptual quality of an audio signal. The goal of the digital audio restoration field [1] is, ideally, to improve the quality of audio signals extracted from old recordings, such as wax cylinders, 78 rpm, longplaying records, magnetic tape, and even digital media matrices. The usual approach consists of finding the best way to capture and transfer the recorded sound from the original matrices to a digital medium and, after that, applying digital signal-processing techniques to remove any disturbance or noise produced by the recording and reproducing system. The most common tasks of audio restoration algorithms are to remove impulsive noise and reduce broad-band noise from the degraded audio sources. Whereas localized disturbances, at least those of short duration, are relatively easy to treat, dealing with global types of degradation is still a challenging task. In particular, in the broad-band noise-reduction problem the goal is to find better tradeoffs between effective noise reduction and signal distortion [1], [2]. Although the perceptual quality of the restored signals plays an important role in this matter, only recently have psychoacoustic criteria been proposed for audio enhancement purposes [3], [4], still bounded by the lack of an observable clean reference signal. Usually audio restoration algorithms employ signal modeling techniques, which deal with the information available in the surface presentation of audio signals, that is, the attempt to model the waveform representation of the audio signal. In sound source modeling (SSM) techniques, however, the goal is to model the phenomenon that has generated the waveform. As a natural consequence, a structured audio representation [5] is required in SSM. In addition to SSM, models for the propagation medium and the receptor characteristics have been increasingly employed in audio signal processing. In [6] a general framework for audio and musical signal processing is described. It shows the hierarchical scales and relationships among several levels of audio representations. In fact, actual challenging audio signal-processing applications seem to move toward the incorporation of higher representation levels of audio signals, such as the objectand content-based ones. Among those applications it is possible to cite sound source recognition [7], sound source separation [8], music retrieval, automatic transcription of music [9], object-based sound source modeling [10], and sound synthesis [11]. Due to the requirement of a structured audio representation when using SSM, its practical use for the analysis and synthesis of audio signals is still limited to specific cases. It is easy to see that for general cases, such as analysis and synthesis of polyphonic music, the SSM-based system faces difficult tasks. The analysis part requires