Prototype-Rectified Iterative Self-supervised Manifold Denoising under Severe Acoustic Shift

DiTAR+: Dual Optimization for Robust Autoregressive Diffusion Speech Synthesis

Continuous-latent Autoregressive Diffusion Transformer (AR-DiT) models have demonstrated immense potential in zero-shot speech generation. However, they still suffer from limited decoding stability when synthesizing long utterances or complex linguistic structures. This instability primarily stems from a restricted historical receptive field and an acoustic inertia dependency within the diffusion decoder, which causes the model to ignore semantic conditions. To address these challenges, we propose DiTAR+, a dual-optimization framework. First, we introduce Dilated Context Sampling to expand the macro-level historical receptive field without violating physical temporal continuity, thereby preventing cumulative error propagation. Second, we propose Hierarchical Acoustic Masking to prevent shallow layers from attending to acoustic pre-context, explicitly decoupling semantic alignment from acoustic detail reconstruction. Extensive experiments show that our framework effectively mitigates pronunciation errors and semantic hallucinations, enhances generation robustness on challenging sentences, and maintains exceptionally high speaker similarity throughout the entirety of long-form utterances. On the linguistically challenging ZH-Hard set, DiTAR+ reduces the word error rate from 12.478% to 9.893%, and on extended utterances of 25 to 35 seconds it improves speaker similarity from 0.741 to 0.759 while simultaneously lowering the word error rate from 2.778% to 2.173%, outperforming both discrete-token and pure flow-matching baselines.

Mind the Microphone Gap: Benchmarking Array Upsampling Strategies for Latent Acoustic Mapping

Results show that the original full-resolution LAM is the strongest, that separately trained lightweight models are the most competitive learned approaches, and that representation alignment between the upsampler and LAM matters more than model complexity.

Self-supervised audio representation learning model based on time-frequency decoupling and masked reconstruction

SPECTRA: Subspace-Preserving Embedding Calibration, Transport, and Replay for Fully Few-Shot Class-Incremental Audio Classification

Fully few-shot class-incremental audio classification (FFCAC) requires recognizing new sound classes from only a handful of labeled examples per session, without forgetting previously learned classes and without any large base dataset. Existing methods typically freeze a pre-trained audio--language encoder and classify with point prototypes, but they suffer from significant performance degradation throughout the sessions due to generic feature representations. We propose SPECTRA, a framework built on a frozen encoder which adds three components. (i) a lightweight trainable adapter that calibrates the generic embeddings to the task; (ii) subspace feature replay, an exemplar-free anti-forgetting scheme that replays old classes by sampling from the low-rank subspace of their stored features; and (iii) a transductive optimal-transport refinement of prototypes at test time. Our central finding is that the subspace structure of the replay diminishes forgetting and outperforms naive Gaussian replay of equal variance. On three FFCAC benchmarks (NSynth-100, FSC-89, LS-100), SPECTRA improves average accuracy and reduces forgetting over current state-of-the-art methods, and our ablations statistically validate each component.

Beyond Reconstruction: Full-Context Generative DiT for Music Generation

FullDiT is introduced, a conditional DiT that fuses eight frame-aligned RVQ streams with independently encoded captions and lyrics and uses non-causal self-attention over the complete acoustic latent sequence and outperforms five commercial systems on 15 of 18 automatic metrics.

Blind Device-Response Interpolation for Intermediate-Domain Unsupervised Adaptation in Cross-Device Acoustic Scene Classification

Recording-device mismatch is one of the dominant failure modes of acoustic scene classification: a model trained on one microphone degrades sharply on an unseen one. Aligning source and target in a single adversarial step is unstable when the device gap is large, and recent intermediate-domain formulations mitigate this by inserting bridge domains, but they build those bridges from hand-specified parametric device operators that are unavailable at deployment time. We propose BRIDA, an intermediate-domain unsupervised adaptation framework whose bridges are derived from the device response itself. A blind estimator recovers the relative log-mel magnitude response between the labelled source devices and the unlabelled target device from long-term spectral statistics alone, with an optional class-balanced refinement for the case where the two unlabelled pools differ in scene composition. Because a magnitude filter is additive in the log-mel domain, fractional powers of the estimate yield a continuum of label-preserving bridge domains at essentially zero cost; we traverse that continuum with a curriculum and regularise it with an ordinal path-position adversary, a residual maximum-mean-discrepancy anchor and confidence-gated target prototypes. Experiments use 1648 real recordings and 66 measured microphone impulse responses under a leave-one-device-out protocol with disjoint source, target and evaluation content. Averaged over 3 unseen devices and 3 seeds, BRIDA attains 53.00 ± 4.29% target accuracy and 52.99 ± 4.58% macro-F1, improving on source-only training by 7.83 points and on a discrete two-bridge intermediate-domain baseline by 0.61 points, with no inference-time overhead.