DJ mixing requires coordinating multiple musical attributes, yet existing approaches in music information retrieval lack a unified framework for modeling these factors. In this work, we propose MixWeaver, an interpretable framework for automatic DJ transition that captures key aspects of transition quality, including musical structure and alignment, stem-wise coordination, transition timing, and perceptual consistency. We formulate DJ transition as an optimization problem over stem-wise gain functions, enabling a principled representation of how musical elements are exchanged between two tracks over time. Our method constructs a candidate space by combining stem-specific transition templates and musically aligned windows for four stems, and evaluates each candidate using an interpretable objective based on stem-overlap control, gain smoothness, and total energy stability. A genetic algorithm is used to search this large combinatorial space efficiently. Experiments on more than 2,000 transition pairs show that the full system consistently outperforms candidate-space and scoring ablations in objective evaluation, and a subjective listening study further shows preference over a rule-based baseline. These results highlight the value of interpretable stem-aware design for automatic DJ transition generation.

1.In response to Reviewer1 (Weakness1), Reviewers2 and 3, and the Meta-Reviewer, we expanded the discussion of DJtransGAN, clarified its differences from our training-free stem-level approach, explained our selection of MosaikBox, and acknowledged the missing direct comparison as a limitation. 2.In response to Reviewer1 (Weakness2), we clarify that the current objective targets smooth intro/outro blending. Hard cuts and abrupt transitions may be unfairly penalized by the smoothness terms, so support for mode-specific transition objectives is left for future work. 3.In response to Reviewer1 (Weakness3), we clarified “basic_swap_only” and removed the ITU-R BS.1534-3 citation because our study did not use reference or anchor conditions and was not a MUSHRA evaluation. 4.In response to Reviewer1 (Weakness5), we removed RQ4 on objective-subjective consistency because 10 transition pairs are insufficient for a reliable correlation analysis. 5.In response to Reviewer2. We’ve defined transition generation in the main text. We didn’t use DJ mixing because it involves scratching or fancy sound effects, which are not implemented here, so “transition” will be more proper; EQ is a conventional control in DJ mixing, while stem separation provides a more fine-grained way to manipulate semantic components such as drums, bass, and vocals. Jointly optimizing stem-level gains and EQ parameters would enlarge the control and search space, making the problem non-trivial. We therefore leave this extension for future work, EQ-related spectral compatibility could be incorporated directly into the scoring function; The dataset was compiled from the authors’ personal music collections and is not a public dataset. We have clarified its source in the revised paper; Changed all “et al.” to “\emph{et al.}”. 6.We corrected Eqs.(9) and (13) to match the implementation, and added the original BeatNet reference alongside the existing citation to the underlying joint beat and downbeat tracking method.