The Application of Neural Networks in Guitar Style Transformation

Abstract: This paper explores techniques in audio style transfer using deep learning models and spectral analysis. Our approach transforms raw audio signals into the frequency domain using Short-Time Fourier Transform (STFT), facilitating the extraction of spatial features crucial for audio synthesis. Central to our methodology is utilizing neural networks such as RandomCNN and EfficientNetV2-S for effective feature extraction from spectrograms. These networks are adapted to preserve essential spectral characteristics while enhancing computational efficiency. Content loss functions ensure fidelity between generated and source audio, quantified through mean square error in feature space. Style transfer is achieved via Gram matrices, capturing temporal correlations to imbue generated audio with stylistic attributes akin to reference audio. For reconstructing audio from spectrograms, we employ the Fast Griffin-Lim Algorithm (FGLA), which iteratively estimates phase information from magnitude data to produce high-quality audio outputs. Experimental validation using the Guitarset dataset demonstrates the efficacy of our approach across diverse musical genres, including Funk and Bossa Nova. Spectrographic analysis of generated outputs validates the preservation of content integrity and faithful adoption of stylistic elements from reference audio. This research investigates the capabilities of neural networks in audio synthesis, promising applications in music production and artistic expression through nuanced style fusion and high-fidelity audio reproduction.