Inside the Latent Flow: Causal Deciphering of Attention Dynamics in Audio Separation Foundation Models
Authors: Yuxuan Chen, Haoyuan Yu, Peize He
Organizations: Jilin University · Hunan University · University of Electronic Science and Technology of China
Abstract
Flow-matching transformers achieve strong audio separation, yet their attention dynamics are opaque. We adapt established causal-intervention principles into a deterministic, inference-time probing protocol for SAM Audio. Orthogonal probing uncovers a dual-pathway text-conditioning mechanism: additive injections control semantic identity, while cross-attention refines acoustic structure. We observe an asynchronous layerwise convergence: stable layers build temporal scaffolds early, whereas fast layers continue resolving artifacts during sampling. The model also attenuates temporal segmentation cues to maintain continuous-flow stability. Using these insights, we propose Layer-Selective Attention Caching (LSAC), a training-free acceleration method that caches attention in stable layers. Across acoustic complexities, LSAC cuts self-attention computation by about ~25% with negligible quality loss and yields up to 6.7x higher quality retention than naive step reduction.
Text-guided sound separation enables flexible audio editing, assistive listening, and open-domain source extraction, but systems such as AudioSep remain too expensive for low-latency edge or codec-mediated deployment. Existing neural audio codec separators are efficient, yet largely restricted to fixed stems or closed taxonomies. We introduce CodecSep, a prompt-driven universal sound separation framework that extracts sources directly in neural audio codec latent space. CodecSep combines a frozen DAC backbone with a lightweight FiLM-conditioned Transformer masker driven by CLAP text embeddings, enabling open-vocabulary separation while preserving codec-native efficiency. Across dnr-v2 and five open-domain benchmarks, CodecSep consistently improves over AudioSep in SI-SDR, remains competitive in ViSQOL, and achieves clear gains in human MOS-LQS. Controlled analyses show that fine-grained prompts outperform coarse labels, and that explicit latent masking is substantially more effective than decoder-style latent generation in codec space. Qualitative diagnostics show that neural audio codec latents retain source-dependent structure, which CodecSep exploits mainly through channel-wise source-conditioned modulation. CodecSep also provides a practical code-stream deployment path. When audio is transmitted as neural audio codec codes, CodecSep maps codes to embeddings, separates directly in codec space, and outputs waveforms or re-quantized codes, avoiding the decode-separate-re-encode loop. In this regime, CodecSep requires only 1.35 GMACs end-to-end: about 54 times less compute than AudioSep in the same pipeline and 25 times lower separator-only compute, with much lower latency and memory. More broadly, CodecSep offers a blueprint for codec-native downstream audio processing.
Generative modeling provides a flexible way to model mixture-conditioned source distributions, but iterative diffusion and flow matching models are costly for long music signals. This paper studies joint vocal-accompaniment separation through latent flow matching, where a pretrained variational autoencoder (VAE) maps mixtures and sources into a compact latent space and a flow matching model generates vocal and accompaniment latents jointly. The proposed framework decouples semantic separation from acoustic velocity prediction through a Separation Encoder and a Velocity Decoder. To reduce sampling cost, we further apply latent adversarial post-training inspired by Flow2GAN for few-step generation. Experiments show that latent adversarial refinement can improve perceptual and separation metrics under a reduced sampling budget.
Transformers with global attention capture long-range dependencies but can miss the fine-grained local continuity crucial for speech separation. We propose TF-MossFormer, a time-frequency transformer that combines local and global attention to jointly model short- and long-range contexts for monaural speech separation. At its core is a content-aware sliding-window attention mechanism that dynamically adapts receptive fields for stronger local interactions, avoiding the rigidity of static convolutions. Unlike time-domain chunk-based methods, TF-MossFormer leverages the 2D spectrogram to model structure along both time and frequency axes. Convolutional gating between attention layers further improves feature selection and information flow. TF-MossFormer achieves SI-SDRi of 22.6, 24.0, and 24.4 dB on WSJ0-2Mix with 5.9M, 16.9M, and 25.4M parameters, respectively, outperforming prior approaches.