Audio Style Transfer Using a Modified Lightweight TimbreTron Pipeline

Neural networks have gained extensive usage in various tasks such as classification, object detection, and style transfer, particularly with images and text. Audio style transfer has made progress in several areas, but there is still room for development. This paper presents a lightweight pipeline to solve the audio style transfer problem, modifying the TimbreTron style transfer such that the reduced computational demand allows deployment on edge hardware. Notably, it is difficult to apply common methodologies like accuracy, recall, or precision monitoring while reviewing the findings. Instead, evaluation depends heavily on personal preferences as it is based on subjective auditory perception. The technique that was used to overcome the aforementioned difficulties is examined in the paper. In a survey that was done, most respondents, 54.5%, claimed that the audio quality was identical when switching from Classical to Jazz and back to Classical. Additionally, 36.4% of the participants said that the audio similarity was slightly noticeable.