Machine learning in audio mastering: a comparative study

Machine learning approaches now utilized in audio mastering are transforming traditional workflows. This comparative study examines the effectiveness of supervised and unsupervised methods in the mastering process. Platforms such as LANDR employ supervised models that emulate expert engineers, offering cost-effective options for independent artists, while unsupervised techniques aid spectral balance and dynamic range optimization. The methodology relies on objective metrics—including Distortion Meter, Dynamic Range, Loudness Penalty, Intelligibility, and High Frequency Distortion—along with subjective listening assessments. Statistical analyses show that human engineers surpass AI systems in preserving dynamic range, minimizing distortion, and maintaining sonic clarity, particularly for complex genres like classical and jazz. Empirical research reveals AI mastering causes greater distortion, narrower dynamic range, and higher loudness penalties. In contrast, engineers deliver superior audio quality through broader dynamic range, lower distortion, and enhanced intelligibility. While AI quickly provides reasonable results for simpler styles like Pop and Electronic, human expertise offers advantages for complex compositions where aesthetic judgment is key. These findings indicate that despite technological progress, human know-how remains critically vital in creative decision-making. The study also points to potential for human-machine collaboration in mastering, with AI initially optimizing parameters and engineers making refined aesthetic adjustments to enhance quality. This hybrid approach could unite technological efficiency with artistic excellence. Future work should focus on improving AI's ability to emulate human aesthetic decisions, developing genre-specific mastering, and incorporating techniques like generative adversarial networks to mastering. These advancements may pave the way for hybrid systems fusing human creativity and machine efficiency.