Traditional source separation methods rely on coarse-grained categorical labels through labeling all vocals collectively without distinguishing individual voices in an audio mixture, which inherently limits the ability to isolate single tracks. While fine-grained annotations could partially address this issue, they demand substantial resources and face challenges in extracting tracks from raw signals. To overcome these limitations, we propose to extract each track through decomposing the patterns of data generation. Specifically, we refine Variational Stochastic Dirichlet Process-VAE, a variational autoencoder framework through replacing the standard variational distribution by a variational stochastic Dirichlet process (VSDP). Among our proposed framework, the encoder, leveraging stick-breaking constructions, adaptively partitions the latent space into clusters, while the decoder designed to recover each component achieves implicit signal separation. Its advantage is that the reconstruction target can be shifted from the raw input to its individual components. Experiments demonstrate our method’s efficacy in two scenarios: (1) Under coarse-grained source definitions, it reaches near-state-of-the-art performance (SDR=10.3); (2) For fine-grained track separation, what’s more, it identifies 83% of individual vocal tracks with an average SDR of 7.8, which cannot be obtained by other SOTA methods without the help of annotations.
The coarse-to-fine feature matching paradigm has demonstrated highly effective in point cloud registration. This paradigm progressively propagates feature correspondences from the coarse level to the fine level through hierarchical feature extraction. However, it is limited by the low discriminability of coarse-level features due to insufficient modeling of global geometric structures, which results in unreliable initial correspondences. Furthermore, relying on single-level features leads to the irreversible loss of fine-grained information, especially in low-overlap scenarios. These limitations present significant challenges in maintaining global geometric consistency and result in a high incidence of feature mismatches. To address these limitations, we propose the HFA-Transformer, a novel Hierarchical Feature Aggregation Transformer framework with two key innovations: (1) a feature enhancement mechanism that jointly encodes spatial and channel-wise characteristics of point clouds, enriching the global feature representation; (2) a Hierarchical Feature Aggregation Module that integrates hierarchical features to refine coarse-level correspondence estimation. Extensive experiments conducted on both indoor and outdoor benchmarks validate the superior performance and robustness of the proposed HFA-Transformer.
