This paper tackles the underexplored problem of token sparsification in audio transformers, where self-attention cost grows quadratically with input length. The authors systematically compare pruning and merging strategies across model sizes and datasets, and propose frequency-aware heuristics—Spectral Flatness (SF) and Low-Frequency Energy (LFE)—to score token importance. Their core finding is that audio-specific spectral cues consistently outperform generic attention-only methods in post-training settings, offering a practical path to faster inference on spectrogram-based models.

Text-to-music alignment is usually tested with a single global similarity score that hides which specific musical attributes are missing or misaligned. This paper proposes MQAScore, a fine-grained evaluator that uses a large audio language model to verify each atomic concept—across genre, instrument, mood\/theme, vocal, and usage dimensions—via binary question answering. The work matters because it moves beyond opaque global scores toward interpretable, per-dimension diagnostics that can directly guide improvements in text-to-music generation systems.

This paper proposes FLARE, a framework for learning robust visual speech representations for lipreading by distilling knowledge from Whisper audio models into discrete codes via random-projection quantization, while jointly processing full-face and lip-ROI video streams. The core idea is that richer acoustic supervision and complementary facial dynamics can mitigate the ambiguity and pose sensitivity that plague visual speech recognition. FLARE achieves competitive word error rates on LRS2, LRS3, and the challenging out-of-domain WildVSR benchmark, and the authors back their design with extensive ablations and representation visualizations.

Sound classification with transformers usually demands stacking many layers, which inflates parameters and compute. LAST asks whether the same blocks can be reused to refine only a single class token while keeping audio features fixed, trading depth for recurrence. Their looped Audio Spectrogram Transformer matches or exceeds deeper sequential baselines on AudioSet at roughly half the parameter count, suggesting that how tokens are updated matters as much as how many layers exist.

Speech-based Alzheimer's assessment increasingly relies on self-supervised learning models, but their vulnerability to nuisance acoustic factors remains unclear. This paper asks whether noise and reverberation are merely encoded in SSL representations or actively drive diagnostic predictions. Using controlled input- and representation-space interventions on the ADReSSo dataset, the authors show that acoustic manipulations systematically shift classifier outputs even when no baseline group difference exists, challenging the assumption that predictive accuracy implies robustness.

This paper tackles video-to-speech (V2S) synthesis, where the central challenge is a one-to-many mapping: visually identical lip movements can correspond to different phonemes. The proposed Watch Your Speech (WYS) framework injects explicit textual conditioning—ground-truth text during training and lip-reader predictions at inference—into a conditional flow matching model via an attention-based fusion module. The key question is whether training-time textual guidance can outperform inference-only classifier guidance while preserving naturalness.

This paper tackles the problem of speaker traits entangling with synthesis source information in speech deepfake source verification. The authors propose a Speaker-Disentangled Metric Learning (SDML) framework that combines Chebyshev polynomial approximations for gradient stability with Riemannian geometry (hyperbolic space) to separate speaker identity from source generator artifacts. Evaluated on four new cross-protocols using the MLAAD benchmark, the method aims to prevent models from relying on speaker shortcuts when verifying synthetic speech origins.

This paper proposes SqueezeComposer, a long-form music generation framework that tackles computational constraints by applying temporal speed-up (e.g., 2×, 4×, 8×) to compress audio sequences before generation. The core idea is to generate music in an accelerated domain using diffusion models, then restore it to normal speed, theoretically enabling models to produce 10+ minute compositions with fixed memory budgets. The approach is tested on continuation, completion, and singing accompaniment tasks.

This paper studies the coupling between three design axes in audio representation learning: input frontend (raw waveform vs. spectrogram), backbone architecture (Mamba vs. attention), and sequence length. The authors introduce HELIX, a minimal hybrid architecture with five bidirectional Mamba layers and one attention bottleneck at matched 8.3M parameter capacity. The key finding is that these choices are not independent: raw waveforms help with Mamba but not attention, attention hurts on short environmental sounds but becomes critical at 30,000 tokens (5 minutes), where pure attention fails with OOM errors and HELIX closes an 11.5-point gap over pure Mamba on speaker identification.

TiCo tackles a critical gap in spoken dialogue models: the inability to control response duration, which is essential for time-constrained scenarios like driving assistants or emergency healthcare. Unlike text length control, speech duration depends on complex factors including phonetics, prosody, and speaking rate. The paper proposes Spoken Time Markers (STMs)—special tokens like <15.0 seconds> inserted during generation—to enable real-time temporal awareness. Using a two-stage post-training framework (self-generated supervised fine-tuning followed by reinforcement learning with verifiable rewards), TiCo equips models to estimate elapsed time and adjust content dynamically to meet target durations.

This paper introduces TaigiSpeech, the first intent recognition dataset for Taiwanese Hokkien—a low-resource language spoken by 65% of Taiwanese elders. With 3,000+ utterances from 21 elderly speakers across emergency and smart-home scenarios, it addresses a critical gap in speech technology for aging populations. The authors also propose keyword-based and audio-visual mining strategies to bootstrap training data from unlabeled video sources.

DiT-Flow tackles multi-condition speech enhancement (noise, reverberation, codec compression) by combining flow matching with a latent Diffusion Transformer (DiT) backbone. The paper proposes operating flow matching in a VAE-compressed latent space for efficiency, introduces StillSonicSet (a synthetic dataset with realistic room acoustics for stationary sources), and applies Mixture-of-LoRA-Experts (MoELoRA) for parameter-efficient adaptation to unseen distortions. The work matters because most SE models fail when deployed on real-world audio with compound distortions unseen during training.