Toward Neural Music Style Transfer: Deep Learning for Automatic Classical-to-Pop Transformation

Musical style transfer seeks to reimagine a piece in a new stylistic domain while preserving its core musical identity, including melody, harmony, and rhythmic structure. Although recent deep learning advances have enabled progress in symbolic composition, timbre modelling, and neural audio generation, style transfer at the full audio level remains highly challenging. Existing methods commonly focus on symbolic transformations, isolated timbre conversion, or texture-based spectral manipulation, which limits their ability to achieve coherent genre-level transformation with realistic production characteristics. In particular, converting classical music into contemporary pop style requires simultaneously reshaping timbre, articulation, ornamentation, expressive dynamics, and modern production aesthetics without compromising musical fidelity. To address this problem, we present a two-stage deep learning framework for end-to-end classical-to-pop audio style transfer. The system first learns a genre-conditioned latent representation that captures multi-dimensional stylistic cues, and then integrates these features into a content-preserving transformation network for waveform synthesis. Harmonic–percussive decomposition, content-aligned losses, and a high-fidelity neural vocoder are incorporated to enhance clarity, rhythmic articulation, and expressive nuance. Across objective and subjective evaluations, our approach outperforms state-of-the-art symbolic, adversarial, and diffusion-based baselines in content preservation, style fidelity, perceptual quality, and Fréchet Audio Distance. Listening tests confirm that the transformed outputs maintain classical musical identity while rendering convincing pop timbre and production polish, establishing a strong foundation for creative music generation and cross-genre audio synthesis.