The second section of Obsession lasts from 3:08 to 5:04 in the recording of the entire composition. While the first section deals with obscuring the actual guitar, supplanting digital effects on fragments featuring characteristic instrumental techniques or juxtaposing real and artificial feedback sounds, the main feature of Section 2 is the use of software to explore hybridization. Here, the sound of the recorded acoustic instrument and its synthetic counterparts are computationally analyzed, broken apart, and recombined. This is achieved by layering material resulting from cross-synthesis, a technique for digitally blending sounds that originated in the 1980s, along with an AI-based approach for encoding prominent timbre information from audio files. These recent algorithmic possibilities highlight new intertextual connections for musical instrument sounds as compositional drivers in electroacoustic music.
Cross-synthesis involves reinforcing the acoustic frequencies that are most present in multiple sound sources and diminishing weaker ones. The software breaks the entire spectrum of frequencies present in each sound into equally sized groups (these are called frequency bins) and then reassembles the sound using the louder version of each group. While more sophisticated methods for using software to explore timbre do exist, I decided to use this technique for two reasons. First, to establish an aural reference point to an earlier period in electroacoustic music. Despite being used to hybridize any number of sounds, cross-synthesis can produce unique sonic artefacts that are by-products of computational processes. Second, to use this electronic sound residue as part of the data used in more recent algorithmic approaches for exploring timbre that also play a role in the composition. Cross-synthesis of real guitar recordings with FM guitar presets, playing the same technique, emphasizes the subjectivity involved in trying to hybridize sounds, as there are various ways to mix the acoustic components of sounds – some of which might be scientifically accurate but come with unexpected results that do not feel authentically related to the source.
Developed at IRCAM (Institut de Recherche et Coordination Acoustique/Musique), RAVE is a music technology project that encodes the timbre by using deep learning to train a model based on, ideally, hours of audio recordings. This process produces a file that essentially abstracts the more prominent timbral features of the training material so that a sound source received by music software (e.g., an instrument or recording in a digital audio workstation) can be transformed to assume sonic features of the training audio. The audio recordings that I made for my RAVE AI model for Obsession consisted of both conventional melodic sounds played on the acoustic guitar, along with unaltered versions of the tapping and noise sounds used as raw material during Section 1 of the piece. Once trained, the plugin was used to transform the various artificial guitar sources, such as The Synthesis Toolkit instruments and my own custom-built FM8 presets, to take on characteristics of the actual acoustic guitar used for various tapping and strumming effects.
Theorist Erin Manning's analysis of Jim Campbell's visual installation series Motion and Rest supports the role of the viewer and shows how ambiguity between the virtual and the actual does not detract but, in fact, contributes to feelings of living presence. Campbell's use of plexiglass (as a transducer of light) and hundreds of LEDs makes the spatial position of the visitor a crucial factor in the production of the work. This can result in various stages between a pixelated and an actual representation of a disabled person walking. Manning’s discussion of Motion and Rest presents a feedback loop where analog images are digitized and digital pixels are activated in a kind of re-becoming analog (Manning 2008: 331). The following example shows the use of RAVE to transform strummed guitar chords made with FM synthesis to resemble the model trained on melodic and inharmonic noise sounds from an acoustic guitar. The encoding and decoding of the training data, which in a sense represents a form of machine listening, pushes the artificiality of the computer-generated plucks and strums towards various prominent timbre features of the actual guitars used as training audio, presenting certain aspects of the analog recorded guitar sounds while temporally disorganized in a way that reveals their digital origin.
In a sense, using composition to explore instrument synthesis as used throughout other electroacoustic works is not about surveying the accuracy of a particular technique. With feedback or plectrum noises, the instrument is so associated with technological intervention that audio artefacts from digital realizations can add a new layer of performativity to the sound source. This pursuit of latent digital noises becomes evident with the use of the RAVE plugin during the opening of Section 2, where input sounds from synthesis match the timbre of the acoustic guitar training audio (harmonics, low drone sounds from detuning strings, tapping effects, etc). Though the modeling is in some ways more accurate than earlier designs using techniques such as FM synthesis, the way in which the software rapidly transitions to match different aspects of the training material creates a highly disjunct rhythmic complexity that is not idiomatic of any style of guitar playing. This can be thought of as an alternative way of dealing with sonic referentiality in acousmatic music. While source abstraction is not present through dense layering of sounds or the use of unusual microphone placements, it is achieved by making familiar sounds take on unfamiliar temporal behaviors.