Musical mashups increasingly serve as a testbed for computational creativity, yet existing automated systems often reduce harmonic compatibility to surface-level matches in key or chord labels. This paper proposes a two-phase framework for chord latent decoupling that treats harmony as a transferable and disentangled latent dimension, enabling controllable mashup generation beyond traditional signal-level heuristics. Our method extracts chord latents with timbre-invariant training, and adversarial objectives to factorize music into (i) a harmony latent capturing chordal progression, and (ii) a complementary non-harmony latent encoding other information like melody, rhythm, and timbre. A joint decoder reconstructs coherent audio from these disentangled components, permitting crosssong recomposition where one track's harmonic structure guides another's musical content. Experiments on large-scale music datasets, combined with perceptual listening studies, demonstrate that proximity in the learned harmony space correlates with perceived mashup compatibility, and that harmonic transfer produces coherent, musically interpretable results. We discuss limitations in supervision granularity, residual leakage, and decoder fidelity, and outline some salient directions: scalable pseudo-label distillation, cross-modal harmony embeddings for retrieval and recommendation, and structure-aware alignment for long-horizon generation. Our results highlight chord-latent decoupling as a data-driven path to interpretable control in AI music generation.

Music mashups integrate elements from different songs to create surprising and engaging listening experiences. Typically, a mashup combines the vocal track of a base song with the instrumental tracks of complementary songs. Automating the production of mashups has been an area of research for decades. Traditional approaches utilize rule-based methods, such as matching tempo and harmonic similarity, to select optimal segments for mashup generation. More recent techniques leverage neural networks to classify segment compatibility. However, both approaches primarily focus on layering segments that are generally compatible, without ensuring their detailed integration and alignment with the vocal track. In contrast, we introduce a novel approach using Graph Neural Networks (GNNs) that learns to rearrange instrumental segments to better align with the base vocal track, resulting in more surprising and accurate music blends for mashup generation. Additionally, we conducted subjective listening tests to evaluate our generated mashups against a baseline model using the same song pairs and the original base songs, assessing our model’s performance. Generated mashups used for evaluation can be found in https://anonymousmus.github.io/bnr.github.io/.

Music mixing is a process that involves fine-tuning the levels, dynamics, and frequency content of musical elements to ensure clarity and harmony in the final music production. In this paper, we present an automatic mixing system based on diffusion models to correct imbalances in music mixes. We manipulate the well mixed stems’ short-time Fourier transform randomly to simulate the frequency and level imbalances commonly encountered in real-world scenarios. The difference between the imbalanced mix and the original mix is treated as noise for the diffusion model to predict, enabling the reverse denoising process to generate an automated mix. We evaluate our model’s performance by calculating the signal-to-distortion ratio between the original and predicted mixes. These results are compared with baseline automatic mixing models, demonstrating significant improvements. Test set results in audio: https://aimg2025submission.github.io/diffmusicmix/

Time-sync video tagging aims to automatically generate tags for each video shot. It can improve the user's experience in previewing a video's timeline structure compared to traditional schemes that tag an entire video clip. In this paper, we propose a new application which extracts time-sync video tags by automatically exploiting crowdsourced comments from video websites such as Nico Nico Douga, where videos are commented on by online crowd users in a time-sync manner. The challenge of the proposed application is that users with bias interact with one another frequently and bring noise into the data, while the comments are too sparse to compensate for the noise. Previous techniques are unable to handle this task well as they consider video semantics independently, which may overfit the sparse comments in each shot and thus fail to provide accurate modeling. To resolve these issues, we propose a novel temporal and personalized topic model that jointly considers temporal dependencies between video semantics, users' interaction in commenting, and users' preferences as prior knowledge. Our proposed model shares knowledge across video shots via users to enrich the short comments, and peels off user interaction and user bias to solve the noisy-comment problem. Log-likelihood analyses and user studies on large datasets show that the proposed model outperforms several state-of-the-art baselines in video tagging quality. Case studies also demonstrate our model's capability of extracting tags from the crowdsourced short and noisy comments.

Music emotion recognition (MER) aims to recognize the affective content of a piece of music, which is important for applications such as automatic soundtrack generation and music recommendation. MER is commonly formulated as a supervised learning problem. In practice, except for Pop music, there is little labeled data in most genres. In addition, emotion is genre specific in music and thus the labeled data of Pop music cannot be used for other genres. In this paper, we aim to solve the genre-specific MER problem by exploiting two kinds of auxiliary data: unlabeled songs and social tags. However, using these two kinds of data effectively is a non-trivial task, e.g. tags are noisy and therefore cannot be treated as fully trustworthy. To build an accurate model with the help from the unlabeled songs and noisy tags, we present SMART, which stands for Semi-Supervised Music Affective Emotion Recognition with Social Tagging, combining of a graph-based semi-supervised learning algorithm with a novel tag refinement method. Experiments on the Million Song Dataset show that our proposed approach, trained with only 10 labeled instances, is as accurate as Support Vector Regression trained with 750 labeled songs.