We present a systematic study of fine-tuning OpenAI's Whisper large-v3 for Swiss German ASR, using 1,367 hours of broadcast speech paired with Standard German subtitles as weak supervision. Through 16 iterative training runs on an NVIDIA DGX Spark (Grace Blackwell, 128 GB unified memory, up to 1 PFLOP FP4), we compare LoRA and full fine-tuning of the 1.55B-parameter model, investigate hallucination root causes, and quantify the effect of data quality, subtitle alignment, and training strategy. Our best model achieves 25.6% measured WER on the All Swiss German Dialects Test Set (ASGDTS) in an honest evaluation on strictly disjoint data. A harmonized error analysis separating genuine errors from valid stylistic variation (tense, word order, Swiss orthography) yields a content WER (cWER) of 13.8%, counting only actual recognition failures. Bias-corrected estimation reduces this to 8.5%, suggesting the true error rate is roughly one third of measured WER. We demonstrate that published state-of-the-art Swiss German ASR results (17.1-17.5% WER) are inflated by benchmark contamination: a vanilla Whisper model self-trained on the ASGDTS test set with zero Swiss German data achieves 13.88% WER, surpassing all published systems. Experiments with Phi-4-multimodal show an even stronger memorization effect (3.9% WER), revealing that the benchmark primarily measures convention matching rather than dialectal comprehension. We release two models, a LoRA adapter (25.32% WER, 13.9% cWER) and a full fine-tuned model (25.60% WER, 13.8% cWER), among the few publicly available, honestly evaluated Whisper models for Swiss German, under Apache 2.0 with full reproducibility, requiring no institutional data agreements.

Automated stuttering detection (ASD) systems struggle with paediatric speech due to high acoustic variability in developing voices and the subtle distinction between pathological stuttering and typical developmental disfluencies. We introduce Paediatric-HGNN, a framework using a Context-aware Part-whole Interaction Network (CaPIN) tailored for paediatric data. Instead of conventional 1D signal modelling, our approach builds a heterogeneous graph capturing hierarchical relationships between lexical units (word nodes) and fine-grained acoustic segments (frame nodes). Trained on curated paediatric corpora (UCLASS and FluencyBank), Paediatric-HGNN achieves 82.4% weighted accuracy and a Typical Disfluency F1-score of 0.386. Modelling hierarchical lexical-acoustic interactions captures developmental "searching" behaviour, offering a more robust and interpretable tool for early clinical intervention.

Multilingual ASR models such as Whisper perform well on high-resource languages but exhibit substantially higher Word Error Rates (WER) for Dravidian languages compared to Indo-Aryan ones. Through linguistic and dataset analysis, we show that Dravidian languages have longer words, higher vocabulary diversity, and lower repetition, resulting in sparse token distributions and frequent character-level substitution errors. Baseline fine-tuning further reveals decoder imbalance between self-attention (linguistic context) and cross-attention (acoustic cues). Although synthetic token-repetition experiments indicate potential gains, they are impractical. Motivated by these observations, we introduce two decoder-level enhancements: Weighted-Attention, which adaptively balances attention sources, and Self-Conditioning, which reinjects intermediate predictions to improve token consistency. Experiments demonstrate consistent WER reductions for low-resource and agglutinative languages.

Automatic speech recognition systems often produce confident yet incorrect transcriptions under noisy or ambiguous conditions, which can be misleading for both users and downstream applications. Standard evaluation based on Word Error Rate focuses solely on accuracy and fails to capture transcription reliability. We introduce an abstention-aware transcription framework that enables ASR models to explicitly abstain from uncertain segments. To evaluate reliability under abstention, we propose RAS, a reliability-oriented metric that balances transcription informativeness and error aversion, with its trade-off parameter calibrated by human preference. We then train an abstention-aware ASR model through supervised bootstrapping followed by reinforcement learning. Our experiments demonstrate substantial improvements in transcription reliability while maintaining competitive accuracy.

We present a voice conversion (VC) framework that utilizes K-Nearest Neighbors (KNN) retrieval over WavLM representations to align non-parallel source and target speech, constructing synthetic training pairs for supervised learning. The retrieved segments serve as synthetic inputs, while real target audio provides ground-truth outputs, forming a synthetic-to-real training paradigm that naturally supports multilingual data without requiring parallel corpora or explicit alignment. To ensure consistent target-speaker identity, we incorporate a speaker loss derived from a pretrained speaker verification model. Experiments across multiple languages demonstrate that the proposed approach achieves high naturalness and strong speaker similarity, outperforming competitive VC baselines, despite being trained exclusively on English data. Samples can be accessed at: https://palindromic-vc.github.io.

Codec-based autoregressive (AR) speech language models have achieved strong text-to-speech (TTS) quality by modeling speech as sequences of discrete audio tokens with large pretrained backbones. However, this token-level formulation creates a structural efficiency bottleneck: speech-token sequences are much longer than text sequences, requiring the AR backbone to perform causal computation at every token position and maintain a KV cache that grows with the sequence length. We introduce TLDR, a patch-based autoregressive framework that accelerates codec-based AR-TTS by shifting the causal modeling from token-level speech sequences to patch-level sequences. TLDR groups consecutive codec tokens into compact latent patches using a lightweight compressor, models the resulting shorter patch sequence with a frozen pretrained AR-TTS backbone adapted by LoRA, and reconstructs fine-grained speech tokens within each patch using a speaker-conditioned extractor. With a patch size of 4, TLDR achieves a 1.8x inference speedup over the baseline AR-TTS model and reduces global KV-cache memory by up to 75%. Experimental results indicate that patch-level global causal modeling can be a practical way to reduce the inference cost of pretrained codec-based AR-TTS systems without replacing the existing modules.

Recent state-of-the-art (SOTA) text-to-speech (TTS) systems typically adopt a cascaded pipeline consisting of a speech tokenizer, an autoregressive large language model (LLM), and a diffusion based flow-matching (FM) model, with these components trained independently. In this paper, we propose a fully end-to-end (E2E) optimization framework that unifies the training of the speech tokenizer, LLM, FM model, and an additional reward model (RM). Specifically, we first jointly optimize the tokenizer using multi-task objectives derived from reconstruction for FM, next-token prediction for LLM, and multi recognition task for RM. This joint training encourages the discrete speech token space to capture acoustically and semantically salient information that is better tailored to TTS. We then further optimize the LLM using downstream reconstruction and recognition by FM and RM, which reduces inference-time mismatch and steers the LLM toward more preferred generations. Experimental results show that our E2E framework consistently outperforms cascaded baselines. On the Seed-TTS-Eval benchmark, our system achieves a word error rate (WER) of 0.78% and 1.56%, a new SOTA result with a 0.6B-parameter LLM and 0.5B-parameter FM model. These results validate that holistic E2E optimization is critical for improving discrete-token-based TTS systems with a much simpler training pipeline.

Removing intermediate representations and separately trained decoding stages has become an important direction in generative modeling. In text-to-speech, however, high-quality systems are still commonly built through an intermediate acoustic representation before waveform synthesis. In this work, we present BareWave, a fully waveform-native framework for direct text-to-wave generation in flow-matching TTS. We consider this setting to raise three training challenges: raw-waveform modeling lacks a strong pretrained representational scaffold, different stages of training benefit from different noise schedules, and data-space perceptual objectives do not automatically share the temporal structure of the velocity-space flow objective. As a result, direct waveform training is hard to optimize efficiently, hard to push toward a strong final operating point with a fixed recipe, and hard to integrate effective perceptual refinement. Guided by this view, we develop a direct text-to-wave training framework that combines training-time representation alignment, staged noise scheduling, and velocity-aware perceptual alignment (VAPA), while preserving a single waveform-native inference path without pretrained components at test time. Experiments on zero-shot voice cloning show that strong intelligibility, speaker similarity, and naturalness can be achieved under a fully waveform-native inference path, supporting waveform-native flow-matching TTS as a practical direction. Project page with audio demos is available at https://barewave.github.io/.

Streaming zero-shot voice conversion (VC) has become increasingly popular due to its potential for real-time applications. The recently proposed MeanVC achieves lightweight streaming zero-shot VC, but it has several limitations: its chunk-wise autoregressive denoising doubles the effective training sequence length, conversion quality degrades under small-chunk settings, and its timbre encoder directly relies on reference mel-spectrograms, making it sensitive to reference audio quality. To address these limitations we propose MeanVC 2. We introduce future-receptive chunking (FRC), which explicitly schedules past and future receptive fields across diffusion transformer decoder layers and removes clean-chunk teacher forcing. By incorporating bounded future context, FRC enables stable conversion with a 40 ms chunk size. We further introduce a universal timbre token encoder, which constructs a timbre representation from a global speaker embedding and retrieves fine-grained timbre cues via cross-attention, improving robustness to low-quality references and enhancing zero-shot speaker similarity. Experimental results show that MeanVC 2 significantly outperforms MeanVC, while reducing latency from 211 ms to 110 ms. Audio samples are publicly available. The source code will be publicly released.

Speech restoration through silent speech interfaces (SSIs) has emerged as a promising assistive technology for individuals with impaired or absent laryngeal voice production. Among non-invasive SSI modalities, surface electromyography (sEMG) and video-based lipreading provide complementary articulatory information, yet their integration for continuous speech synthesis remains underexplored. Moreover, existing multimodal approaches rarely address robustness to modality degradation or temporary sensor failure, limiting their applicability in realistic scenarios. In this work, we propose a masked multimodal speech synthesis framework that jointly leverages sEMG and lipreading signals through modality masking during training. Under multispeaker settings, the proposed approach reduces word error rate by up to 14 absolute percentage points compared to the strongest unimodal baseline. Experimental results not only show that masking strategies are critical for these performance gains and robustness under low-bitrate conditions, but also that they generalize better than degradation-specific data augmentations in the presence of modality absence conditions. Phone-level analyses further reveal complementary contributions across modalities, with particularly strong benefits for vowels and for specific consonant groups. Overall, these findings demonstrate the effectiveness and robustness of masked multimodal integration for silent speech synthesis, although adaptation to laryngectomized speakers remains an open research challenge.

Large language models (LLMs) have demonstrated promising performance in both automatic speech recognition (ASR) and text-to-speech (TTS) systems, gradually becoming the mainstream approach. However, most current approaches address these tasks separately rather than through a unified framework. This work aims to integrate these two tasks into one unified model. Although discrete speech tokenization enables joint modeling, its inherent information loss limits performance in both recognition and generation. In this work, we present UniVoice, a unified LLM framework through continuous representations that seamlessly integrates speech recognition and synthesis within a single model. Our approach combines the strengths of autoregressive modeling for speech recognition with flow matching for high-quality generation. To mitigate the inherent divergence between autoregressive and flow-matching models, we further design a dual attention mechanism, which switches between a causal mask for recognition and a bidirectional attention mask for synthesis. Furthermore, the proposed text-prefix-conditioned speech infilling method enables high-fidelity zero-shot voice cloning. Experimental results demonstrate that our method can achieve or exceed current single-task modeling methods in both ASR and zero-shot TTS tasks. This work explores new possibilities for end-to-end speech understanding and generation. Code is available at https://github.com/gwh22/UniVoice.

We propose Universal Speech Content Factorization (USCF), a simple and invertible linear method for extracting a low-rank speech representation in which speaker timbre is suppressed while phonetic content is preserved. USCF extends Speech Content Factorization, a closed-set voice conversion (VC) method, to an open-set setting by learning a universal speech-to-content mapping via least-squares optimization and deriving speaker-specific transformations from only a few seconds of target speech. We show through embedding analysis that USCF effectively removes speaker-dependent variation. As a zero-shot VC system, USCF achieves competitive intelligibility, naturalness, and speaker similarity compared to methods that require substantially more target-speaker data or additional neural training. Finally, we demonstrate that as a training-efficient timbre-disentangled speech feature, USCF features can serve as the acoustic representation for training timbre-prompted text-to-speech models. Speech samples and code are publicly available.

Many operations on sensory data -- comparison, memory, retrieval, and reasoning -- are naturally expressed over discrete symbolic structures. In language this interface is given by tokens; in audio, it must be learned. Existing audio tokenizers rely on quantization, clustering, or codec reconstruction, assigning tokens locally, so sequence consistency, compactness, length control, termination, and edit similarity are rarely optimized directly. We introduce PairAlign, a framework for compact audio tokenization through sequence-level self-alignment. PairAlign treats tokenization as conditional sequence generation: an encoder maps speech to a continuous condition, and an autoregressive decoder generates tokens from BOS, learning token identity, order, length, and EOS placement. Given two content-preserving views, each view's sequence is trained to be likely under the other's representation, while unrelated examples provide competing sequences. This gives a scalable surrogate for edit-distance preservation while discouraging many-to-one collapse. PairAlign starts from VQ-style tokenization and refines it with EMA-teacher targets, cross-paired teacher forcing, prefix corruption, likelihood contrast, and length control. On 3-second speech, PairAlign learns compact, non-degenerate sequences with broad vocabulary usage and strong cross-view consistency. On retrieval tests, it preserves edit-distance search while reducing archive token count by 55%. A continuous-sweep probe shows lower local overlap than a dense geometric tokenizer, but stronger length control and bounded edit trajectories under 100 ms shifts. PairAlign is a sequence-symbolic predictive learner: like JEPA-style objectives, it predicts an abstract target from another view as a learned variable-length symbolic sequence, not a continuous latent.

Speech tokenizers are a key building block of fully discrete Speech LLMs. Existing tokenizers either prioritize semantic encoding, fuse semantic content with acoustic style inseparably, or achieve incomplete semantic-acoustic disentanglement. To achieve better disentanglement, we propose DSA-Tokenizer, which explicitly disentangles speech into discrete semantic and acoustic tokens via distinct optimization constraints. Specifically, semantic tokens are supervised by ASR to capture linguistic content, while acoustic tokens focus on mel-spectrograms restoration to encode style. We further introduce a hierarchical Flow Matching decoder and a joint reconstruction-context inpainting training strategy, allowing the model to support both high-fidelity reconstruction and cross-utterance voice clone. To speed up inference, we distill the dit decoder to 4-step inference and improve synthesis quality with GAN fine-tuning. Experiments demonstrate that DSA-Tokenizer provides strong semantic-acoustic disentanglement, reliable controllable voice cloning, and efficient high-fidelity generation with low WER/CER. Moreover, our results suggest that disentangled tokenization provides a more effective interface for downstream large-model speech generation. Audio samples are avaialble at https://anonymous.4open.science/w/DSA_Tokenizer_demo/

Although low-bit quantization provides practical means to deploy speaker verification on resource-constrained devices, its effects on speaker verification performance remain poorly understood. In this paper, we study uniform K-means quantization-aware training of ResNet-36 and ResNet-200 through joint layer-wise and score-level analyses. Our layer-wise analysis highlights fragile components and shows that score degradation is not fully explained by weight distortion alone. We identify a clear knee point at 2 bits, with larger score drift and harmful decision flips concentrated near the FP32 threshold. Our score-level analysis reveals where and how score errors emerge under extreme quantization. Building on these findings, we propose a calibrated multi-precision cascade that resolves most trials at 2 bits and escalates only ambiguous cases, achieving performance close to FP32 while preserving the efficiency benefits of low-bit inference with substantially lower compute and memory costs.

Deep-learning speaker verification (SV) increasingly relies on deep neural network backbones, whose environmental impact remains largely undocumented. In this paper, we conduct an evaluation of ResNet architectures trained on VoxCeleb2, varying depth, channel width, and stage distribution, and measure energy consumption and carbon footprint using node-level sensors. Results show a clear point of diminishing returns: deeper or wider models bring only marginal accuracy gains while energy consumption grows steeply. In contrast, mid-sized networks such as ResNet-50 and stage-concentrated variants achieve favorable trade-offs between performance and environmental impact. These findings provide actionable guidelines for designing energy-efficient SV systems.

Speech applications such as meeting transcription and voice agents would benefit from on-device speaker diarization, but practical adoption is limited by inference cost. We study how far a Pyannote 3.1-based pipeline can be accelerated on consumer hardware (an RTX 5070 Ti GPU and an Apple M4 laptop) while preserving diarization error rate (DER). A simple recipe: coarser segmentation stride and per-chunk embedding, yields multi-fold speedups and is DER-neutral on AMI, but degrades sharply on in-the-wild data: on VoxConverse, DER rises from 0.075 to 0.113. We trace the failure to speaker under-counting in the clustering stage, caused by a fixed minimum cluster size interacting with the reduced number of embeddings per speaker. We propose a relative minimum cluster size, mcs = round(f * n) with f = 0.01, which adapts to the embedding budget per recording. A single value of f recovers VoxConverse DER to 0.079 (about 89% of the lost accuracy) while keeping AMI flat, and the accelerated pipeline reaches up to 12.2x speedup on AMI (MPS) over our CAM++ baseline.

Target speech extraction (TSE) typically relies on pre-recorded high-quality enrollment speech, which disrupts user experience and limits feasibility in spontaneous interaction. In this paper, we propose Enroll-on-Wakeup (EoW), a novel framework where the wake-word segment, captured naturally during human-machine interaction, is automatically utilized as the enrollment reference. This eliminates the need for pre-collected speech to enable a seamless experience. We perform the first systematic study of EoW-TSE, evaluating advanced discriminative and generative models under real diverse acoustic conditions. Given the short and noisy nature of wake-word segments, we investigate enrollment augmentation using LLM-based TTS. Results show that while current TSE models face performance degradation in EoW-TSE, TTS-based assistance significantly enhances the listening experience, though gaps remain in speech recognition accuracy.

Using speaker embeddings as conditioning can strengthen speech enhancement, but most methods either require clean enrollment audio or rely on embeddings extracted from noisy speech, which are fragile under noise and domain shift. We propose G-MaP-SE, a guided enhancement framework that builds a clean-speech embedding prior with a Gaussian Mixture Model (GMM) and refines a noisy conditioning embedding by matching it to this prior. The matched prior embedding is then injected into a time-frequency enhancement backbone via a lightweight gated fusion module. Experiments on VoiceBank+DEMAND and DNS Challenge 2020 datasets show that the proposed prior matching consistently outperforms noisy conditioning and substantially narrows the gap to an oracle clean-conditioning upper bound, while requiring no enrollment audio at inference time. The code, audio samples, and checkpoint are available.

Large audio language models (LALMs) are a class of foundation models for audio understanding. Existing LALMs tend to degrade significantly in real-world noisy acoustic conditions where speech and non-speech sounds interfere. While noise-aware fine-tuning can improve robustness, it requires task-specific noisy data and expensive retraining, limiting scalability. To address this issue, we propose Focus-Then-Listen (FTL), a plug-and-play audio enhancer that improves LALMs' noise robustness. Specifically, FTL first separates the input waveform into speech and non-speech, and a modality router is applied to predict the target audio modality (e.g., speech) based on the user's instruction. Finally, a modality-aware fusion block generates a task-adaptive enhanced signal for improved downstream perception and reasoning. Experiments across multiple LALMs and tasks show that FTL improves performance across different noise levels without fine-tuning on LALMs.

We propose Speech Enhancement based on Drifting Models (DriftSE), a novel generative framework that formulates denoising as an equilibrium problem. Rather than relying on iterative sampling, DriftSE natively achieves one-step inference by evolving the pushforward distribution of a mapping function to directly match the clean speech distribution. This evolution is driven by a Drifting Field, a learned correction vector that guides samples toward the high-density regions of the clean distribution, which naturally facilitates training on unpaired data by matching distributions rather than paired samples. We investigate the framework under two formulations: a direct mapping from the noisy observation, and a stochastic conditional generative model from a Gaussian prior. Experiments on the VoiceBank-DEMAND benchmark demonstrate that DriftSE achieves high-fidelity enhancement in a single step, outperforming multi-step diffusion baselines and establishing a new paradigm for speech enhancement.

Automatic speech recognition (ASR) degrades severely in noisy environments. Although speech enhancement (SE) front-ends effectively suppress background noise, they often introduce artifacts that harm recognition. Observation addition (OA) addressed this issue by fusing noisy and SE enhanced speech, improving recognition without modifying the parameters of the SE or ASR models. This paper proposes an intelligibility-guided OA method, where fusion weights are derived from intelligibility estimates obtained directly from the backend ASR. Unlike prior OA methods based on trained neural predictors, the proposed method is training-free, reducing complexity and enhances generalization. Extensive experiments across diverse SE-ASR combinations and datasets demonstrate strong robustness and improvements over existing OA baselines. Additional analyses of intelligibility-guided switching-based alternatives and frame versus utterance-level OA further validate the proposed design.

General speech restoration demands techniques that can interpret complex speech structures under various distortions. While State-Space Models like SEMamba have advanced the state-of-the-art in speech denoising, they are not inherently optimized for critical speech characteristics, such as spectral periodicity or multi-resolution frequency analysis. In this work, we introduce an architecture tailored to incorporate speech-specific features as inductive biases. In particular, we propose the Global, Local, and Periodic (GLP) module, a frequency feature extraction block that effectively and efficiently leverages the properties of frequency bins. Then, we design a multi-resolution parallel time-frequency dual-processing block to capture diverse spectral patterns, and a learnable mapping to further enhance model performance. With all our ideas combined, the proposed SEMamba++ achieves the best performance among multiple baseline models while remaining computationally efficient.

Underwater acoustic classification has a wide array of oceanic applications, but faces challenges due to an increasingly complex acoustic environment. Waveform and spectrogram representations have been primarily used as acoustic data features for classification tasks in this domain. Spectrograms model harmonic dependencies, but these reduced representations can filter out acoustic features relevant for discrimination. While phase information from the waveform allows full characterization of the signal, the original waveform can be noisy and complex, rendering this representation difficult for models to process directly. This paper proposes a dual-encoder neural architecture to simultaneously process acoustic waveforms and spectrograms, leveraging pre-trained backbones and parameter-efficient fine-tuning modules, enabling a domain adaptation. To combine these adapted branches, a novel differentiable fuzzy aggregation mechanism based on the Choquet integral is introduced to balance the temporal and spectral representations. This fusion strategy not only yields higher classification accuracy but also provides interpretability. Specifically, by analyzing the learned fuzzy measures, insights are revealed about class-specific shifts in the network's representation reliance. By dynamically shifting attention to the representation least corrupted by potential asymmetric channel distortions, the proposed gating mechanism mitigates the non-stationary challenges of the underwater environment. Evaluations on the DeepShip and ShipsEar datasets demonstrate that the proposed architecture achieves classification improvements over independent single-encoder baselines, while simultaneously restricting the trainable parameter space. This mitigates the risk of overfitting on limited acoustic datasets while alleviating the computational costs associated with fully fine-tuning foundation models.

Temporal detection problems appear in many fields including time-series estimation, activity recognition and sound event detection (SED). In this work, we propose a new approach to temporal event modeling by explicitly modeling event onsets and offsets, and by introducing boundary-aware optimization and inference strategies that substantially enhance temporal event detection. The presented methodology incorporates new temporal modeling layers - Recurrent Event Detection (RED) and Event Proposal Network (EPN) - which, together with tailored loss functions, enable more effective and precise temporal event detection. We evaluate the proposed method in the SED domain using a subset of the temporally-strongly annotated portion of AudioSet. Experimental results show that our approach not only outperforms traditional frame-wise SED models with state-of-the-art post-processing, but also removes the need for post-processing hyperparameter tuning, and scales to achieve new state-of-the-art performance across all AudioSet Strong classes.

Local density-based score normalization is an effective component of distance-based embedding methods for anomalous sound detection, particularly when data densities vary across conditions or domains. In practice, however, performance depends strongly on neighborhood size. Increasing it can degrade detection accuracy when neighborhood expansion crosses cluster boundaries, violating the locality assumption of local density estimation. This observation motivates adapting the neighborhood size based on locality preservation rather than fixing it in advance. We realize this by proposing cluster exit detection, a lightweight mechanism that identifies distance discontinuities and selects neighborhood sizes accordingly. Experiments across multiple embedding models and datasets show improved robustness to neighborhood-size selection and consistent performance gains.

Symbolic music evaluation for large language models remains fragmented across representations, datasets, and metrics. We introduce LilyBench, a LilyPond-based benchmark that jointly evaluates symbolic music generation and music understanding on the same family of open-weight LLMs. The benchmark includes a 200-prompt generation suite and ten understanding tasks adapted from ABC-Eval, covering syntax, metadata prediction, structural sequencing, and music recognition. Generation quality is evaluated using compile rate, MusPy descriptor distributions via Jensen-Shannon similarity, and LilyBERT-based Fréchet Music Distance (FMD). Experiments on four open-weight models show that executable LilyPond generation is achievable in zero-shot settings, while structural understanding tasks remain challenging despite strong performance on composer and genre recognition. Our experiments also reveal systematic disagreements between descriptor-based and embedding-based metrics, suggesting that symbolic music evaluation benefits from metric triangulation rather than single-score ranking. We release the benchmark, prompt bank, and evaluation code to support future research in symbolic music generation and understanding at https://github.com/CSCPadova/lilybench

The task of music-driven dance generation involves creating coherent dance movements that correspond to the given music. While existing methods can produce physically plausible dances, they often struggle to generalize to out-of-set data. The challenge arises from three aspects: 1) the high diversity of dance movements and significant differences in the distribution of music modalities, which make it difficult to generate music-aligned dance movements. 2) the lack of a large-scale music-dance dataset, which hinders the generation of generalized dance movements from music. 3) The protracted nature of dance movements poses a challenge to the maintenance of a consistent dance style. In this work, we introduce the EnchantDance framework, a state-of-the-art method for dance generation. Due to the redundancy of the original dance sequence along the time axis, EnchantDance first constructs a strong dance latent space and then trains a dance diffusion model on the dance latent space. To address the data gap, we construct a large-scale music-dance dataset, ChoreoSpectrum3D Dataset, which includes four dance genres and has a total duration of 70.32 hours, making it the largest reported music-dance dataset to date. To enhance consistency between music genre and dance style, we pre-train a music genre prediction network using transfer learning and incorporate music genre as extra conditional information in the training of the dance diffusion model. Extensive experiments demonstrate that our proposed framework achieves state-of-the-art performance on dance quality, diversity, and consistency.

Music popularity prediction has attracted growing research interest, with relevance to artists, platforms, and recommendation systems. However, the explosive rise of AI-generated music platforms has created an entirely new and largely unexplored landscape, where a surge of songs is produced and consumed daily without the traditional markers of artist reputation or label backing. Key, yet unexplored in this pursuit is aesthetic quality. We propose APEX, the first large-scale multi-task learning framework for AI-generated music, trained on over 211k songs (10k hours of audio) from Suno and Udio, that jointly predicts engagement-based popularity signals - streams and likes scores - alongside five perceptual aesthetic quality dimensions from frozen audio embeddings extracted from MERT, a self-supervised music understanding model. Aesthetic quality and popularity capture complementary aspects of music that together prove valuable: in an out-of-distribution evaluation on the Music Arena dataset, comprising pairwise human preference battles across eleven generative music systems unseen during training, including aesthetic features consistently improves preference prediction, demonstrating strong generalisation of the learned representations across generative architectures.

Current advancements in Audio Reasoning rely on massive Large Audio-Language Models (LALMs), hindering deployment in resource-constrained environments. We introduce TinyGiantALM, a compact 1.5B efficiency-oriented alternative. Instead of brute-force scaling, we propose an Instruction-Aware Feature Refinement framework using a Query-guided Projector and Semantic Gating to filter acoustic signals based on user intent. On the MMAR benchmark, TinyGiantALM achieves 46.4% zero-shot accuracy, significantly outperforming 7B-13B baselines. While a reasoning gap in logical narrative remains versus 30B+ models and certain trade-offs exist in overly dense or spatial scenes, our approach notably surpasses models up to 8x larger in disentangling mixed-modality environments. These findings demonstrate that architectural precision offers a tangible pathway to secure robust perception capabilities on edge-friendly scales.

Speech-based large language models are typically constrained to spoken replies, which limits their user-facing outputs to what can be verbalized and suppresses text-native capabilities such as code generation, structured analysis, and multi-step reasoning in realtime interaction, for tasks that require persistent, structured, and inspectable intermediate outputs. Existing work improves spoken reasoning or full-duplex turn-taking, but still treats text as a hidden intermediate state or a subordinate modality rather than a first-class output channel. We propose Listen-Write-Speak (LWS), a text-first tri-channel paradigm in which a single autoregressive LLM continuously listens to user audio, writes visible free-form text as its primary output, and speaks a realtime oral response in parallel under a shared causal attention context. This behavior is implemented entirely through a Token Schema, requiring no architectural modifications, and learned via a two-stage data pipeline that synthesizes per-second cognitive annotations consistent with the revealed input timeline. Empirically, LWS demonstrates strong full-duplex interaction on Full-Duplex-Bench, reaches 4.72 on VoiceBench AlpacaEval, achieves 92.6% writing-speaking consistency, and consistently outperforms its internal ablations on URO-Bench. These results suggest that visible writing can serve as a first-class output channel for speech interaction without sacrificing realtime responsiveness. The code and dataset are available on the project page: https://royalzhang.com/project/lws-page/.

Multimodal Large Language Models (MLLMs) effectively integrate text and audio to interpret context in complex interactive dialogues. However, the internal mechanisms by which heterogeneous modalities influence model behavior remain opaque. While Shapley Values (SV) provide a robust, model-agnostic framework for local explainability in text-based NLP, their extension to multimodal data is hindered by cross-channel dependencies, intricate dialogue structures, and the prohibitive computational complexity of dense audio representations. In this work, we formalize a multimodal extension of the Shapley Value framework, treating discrete text tokens and aligned audio segments as cooperative features. To ensure computational feasibility, we deploy a suite of efficient estimation strategies: exact SV computation for low-dimensional inputs and sampling-based approximations - including Monte Carlo permutations and stratified sampling with Neyman-optimal allocation - to minimize variance under constrained computational budgets. To resolve the granularity mismatch between modalities, we propose Spectrogram-Guided Phonetic Alignment (SGPA), a novel preprocessing method that maps high-frequency audio streams to interpretable, word-aligned segments. Our contribution is twofold: first, we provide an open-source, model-agnostic Python package and a companion GUI for the computation and interactive visualization of multimodal attributions. Second, we evaluate our framework using curated subsets of the VoiceBench and Infinity Instruct datasets across diverse multilingual scenarios. Our experimental results reveal that input modality is a primary driver of attribution volatility and demonstrate that standard syntactic importance proxies often fail to predict model attention in multimodal, cross-lingual contexts.

Audio-visual large language models (LLMs) hold strong promise for long-form video understanding, yet their long-video inference is fundamentally limited by the linear growth of video tokens and key-value (KV) caches. We present OmniMem, a memory-efficient streaming framework designed specifically for audio-visual LLMs. Unlike existing compression methods that treat all tokens uniformly, OmniMem introduces a modality-aware memory allocation strategy that separately manages visual and audio contexts, addressing the severe token imbalance between the two modalities. OmniMem further preserves informative and non-redundant KV states through perturbation-aware memory selection, enabling compact memory without sacrificing long-range understanding. To strengthen compression under realistic deployment constraints, we also explore budget-aware fine-tuning, which encourages the model to consolidate useful information into retained memory. Experiments on VideoMME Long, LVBench, and LVOmniBench with video-SALMONN 2+ and Qwen-2.5-Omni show that OmniMem consistently improves over strong training-free compression baselines by 2-4% absolute accuracy under the same memory budgets, with an additional 1-2% gain after fine-tuning.

Audio-Visual Speech Recognition (AVSR) leverages both acoustic and visual information for robust recognition under noise. However, how models balance these modalities remains unclear. We present Dr. SHAP-AV, a framework using Shapley values to analyze modality contributions in AVSR. Through experiments on six models across two benchmarks and varying SNR levels, we introduce three analyses: Global SHAP for overall modality balance, Generative SHAP for contribution dynamics during decoding, and Temporal Alignment SHAP for input-output correspondence. Our findings reveal that models shift toward visual reliance under noise yet maintain high audio contributions even under severe degradation. Modality balance evolves during generation, temporal alignment holds under noise, and SNR is the dominant factor driving modality weighting. These findings expose a persistent audio bias, motivating ad-hoc modality-weighting mechanisms and Shapley-based attribution as a standard AVSR diagnostic.

Synthetic accented speech is a promising way to improve automatic speech recognition (ASR) when real accented recordings are scarce. We ask what makes such data useful for ASR fine-tuning: target-accent phoneme edits that expose the recognizer to accent-specific pronunciations, or random phoneme perturbations that act as augmentation in phoneme space. In a few-shot TTS pipeline, we compare LLM-generated accent edits with matched-rate random substitutions and oracle controls using ground-truth accented phonemes and prosody. Random substitutions recover much of the ASR gain: LLM target-accent edits improve over random by only a small margin, ground-truth phonemes stay close to the random baseline and nearly converge with it as the synthetic ASR fine-tuning set grows larger, and adding ground-truth prosody yields only a modest further gain. Mixing synthetic with real accented speech also stabilizes low-resource fine-tuning, but a fixed synthetic budget can later dilute the information in real data, showing that the real--synthetic ratio matters.

The scale of speech anti-spoofing datasets has grown exponentially over the past decade, driven by the assumption that larger data leads to better performance. However, it remains unclear whether indiscriminate scaling commensurately improves model generalization. This study challenges the "scale-first" paradigm by decoupling the impacts of training data scale versus diversity. Through experiments on representative datasets, we report two key findings: (1) Larger is not always better. Expanding data scale excessively under fixed generation methods yields negligible returns and may even degrade cross-domain generalization due to overfitting.(2) Diversity outweighs scale. A smaller composite training set featuring diverse attacks significantly outperforms larger-scale datasets with limited diversity in cross-dataset evaluations. We conclude that future dataset construction should prioritize the diversity of generation methods over scale to effectively enhance model generalization.

Recent advances in Omni-Multimodal Large Language Models (Omni-MLLMs) have enabled strong integration of vision, audio, and language. However, their audio-visual intelligence (AVI) remains insufficiently evaluated due to the lack of systematic and comprehensive benchmarks. We introduce AVI-Bench, a cognitively inspired benchmark that evaluates Omni-MLLMs across three stages, perception, understanding, and reasoning, through cross-modal tasks requiring joint audio-visual interpretation. This design enables fine-grained diagnosis of model capabilities and failure modes. To further assess robustness beyond familiar domains, we propose AVI-Bench-PriSe, an extension that probes models' primitive audio-visual sensation using unfamiliar, low-semantic stimuli, testing generalization beyond common training distributions. Extensive experiments on both open-source and closed-source models reveal substantial limitations in current Omni-MLLMs. Based on these findings, we present a four-level AVI taxonomy. Overall, AVI-Bench provides a principled evaluation framework to guide the development of more robust and generalizable AVI. Project website: https://fudancvl.github.io/AVI-Bench/

Recent advancements in audio tokenization have significantly enhanced the integration of audio capabilities into large language models (LLMs). However, audio understanding and generation are often treated as distinct tasks, hindering the development of truly unified audio-language models. While instruction tuning has demonstrated remarkable success in improving generalization and zero-shot learning across text and vision, its application to audio remains largely unexplored. A major obstacle is the lack of comprehensive datasets that unify audio understanding and generation. To address this, we introduce Audio-FLAN, a large-scale instruction-tuning dataset covering 80 diverse tasks across speech, music, and sound domains, with over 100 million instances. Audio-FLAN lays the foundation for unified audio-language models that can seamlessly handle both understanding (e.g., transcription, comprehension) and generation (e.g., speech, music, sound) tasks across a wide range of audio domains in a zero-shot manner. The Audio-FLAN dataset is available on HuggingFace and GitHub.

Existing Test-time Adaptation (TTA) studies rely heavily on static and homogeneous corruption protocols, such as ImageNet-C and CIFAR-10-C/100-C, leading to inconsistent evaluation settings and potentially inflated robustness estimates that are compared with real-world situations. TTA lacks a standardized evaluation infrastructure capable of modeling realistic heterogeneous acoustic degradation. We introduce DHAuDS, a standardized benchmark suite for evaluating audio classification TTA robustness under dynamic corruption severity and heterogeneous noise mixtures. Rather than proposing a new TTA algorithm, DHAuDS focuses on exposing robustness limitations that remain hidden under conventional fixed-noise evaluation protocols.

Symbolic music research has relied almost exclusively on MIDI-based datasets; text-based engraving formats such as LilyPond remain unexplored for music understanding. We present BMdataset, a musicologically curated dataset of 393 LilyPond scores (2,646 movements) transcribed by experts directly from original Baroque manuscripts, with metadata covering composer, musical form, instrumentation, and sectional attributes. Building on this resource, we introduce LilyBERT (weights can be found at https://huggingface.co/csc-unipd/lilybert), a CodeBERT-based encoder adapted to symbolic music through vocabulary extension with 115 LilyPond-specific tokens and masked language model pre-training. Linear probing on the out-of-domain Mutopia corpus shows that, despite its modest size (~90M tokens), fine-tuning on BMdataset alone outperforms continuous pre-training on the full PDMX corpus (~15B tokens) for both composer and style classification, demonstrating that small, expertly curated datasets can be more effective than large, noisy corpora for music understanding. Combining broad pre-training with domain-specific fine-tuning yields the best results overall (84.3% composer accuracy), confirming that the two data regimes are complementary. We release the dataset, tokenizer, and model to establish a baseline for representation learning on LilyPond.

Fréchet Audio Distance (FAD) is the de facto standard for evaluating text-to-audio generation, yet its scores depend on the underlying encoder's embedding space. An encoder's training task dictates which acoustic features are preserved or discarded, causing FAD to inherit systematic task-induced biases. We decompose evaluation into Recall, Precision, and Alignment (split into semantic and structural dimensions), using log-scale normalization for fair cross-encoder comparison. Controlled experiments on six encoders across two datasets reveal a four-axis trade-off: reconstruction-based AudioMAE leads precision sensitivity; ASR-trained Whisper dominates structural detection but is blind to signal degradation; classification-trained VGGish maximizes semantic detection but penalizes legitimate intra-class variation. Since no single encoder is a universal evaluator, future metrics must shift toward evaluation-native encoders intrinsically aligned with human perception.

Voice biometric systems face growing threats from spoofing attacks, yet the evaluation of detection models remains inconsistent across datasets. To investigate these unpredictable fluctuations, we conduct a comprehensive benchmark of four self-supervised learning feature extractors paired with four back-end classifiers. We compare the hierarchical local feature extraction of ResNet with the global sequence and relational modeling of attention and graph-based back-ends. Through multi-corpus training across three scenarios and six evaluation datasets, our empirical analysis yields two critical findings. First, we expose a domain bias within the ASVspoof 5 dataset, showing that naive data scaling actively degrades performance. Second, our cross-linguistic analysis reveals that fine-tuning with just 8 hours of target-language data enhances detection robustness. Together, these findings emphasize the critical need for domain-aware and language-specific adaptation in spoofing detection.

Sophisticated generative speech technology can undermined the reliability of voice biometrics. While spoofing detection systems excel when assessed under in-domain conditions, generalisation to out-of-domain settings is often poor. In this paper, we show that such issues could be caused by speaker bias, where models learn individual voice traits rather than markers of manipulation or generation. We propose a teacher-student framework for speaker-invariant spoofing detection that disentangles identity without requiring speaker labels. We leverage a pre-trained speaker recognition teacher to guide a student model via a gradient reversal layer. To control the balance between suppressing cues related to voice identity with the preservation of those related to spoofing detection, we integrate a Variational Information Bottleneck. Evaluations across nine datasets show our model achieves a 25.7% relative reduction to the EER compared to the MHFA baseline.

With the rapid advancement of neural audio codecs, codec-based speech generation (CoSG) systems have become highly powerful. Unfortunately, CoSG also enables the creation of highly realistic deepfake speech, making it easier to mimic an individual's voice and spread misinformation. We refer to this emerging deepfake speech generated by CoSG systems as CodecFake. Detecting such CodecFake is an urgent challenge, yet most existing systems primarily focus on detecting fake speech generated by traditional speech synthesis models. In this paper, we introduce CodecFake+, a large-scale dataset designed to advance CodecFake detection. To our knowledge, CodecFake+ is the largest dataset encompassing the most diverse range of codec architectures. The training set is generated through re-synthesis using 31 publicly available open-source codec models, while the evaluation set includes web-sourced data from 17 advanced CoSG models. We also propose a comprehensive taxonomy that categorizes codecs by their root components: vector quantizer, auxiliary objectives, and decoder types. Our proposed dataset and taxonomy enable detailed analysis at multiple levels to discern the key factors for successful CodecFake detection. At the individual codec level, we validate the effectiveness of using codec re-synthesized speech (CoRS) as training data for large-scale CodecFake detection. At the taxonomy level, we show that detection performance is strongest when the re-synthesis model incorporates disentanglement auxiliary objectives or a frequency-domain decoder. Furthermore, from the perspective of using all the CoRS training data, we show that our proposed taxonomy can be used to select better training data for improving detection performance. Overall, we envision that CodecFake+ will be a valuable resource for both general and fine-grained exploration to develop better anti-spoofing models against CodecFake.

Ambulatory neck-surface acceleration enables non-invasive monitoring of vocal hyperfunction, yet robust biomarkers for its subtypes remain limited. This study investigates the NeckVibe Challenge dataset to distinguish phonotraumatic (PVH) and non-phonotraumatic (NPVH) from healthy controls. We propose a hierarchical feature engineering framework comprising: (i) static, (ii) dynamic, (iii) ratio-based, (iv) coupling features capturing source filter interactions. While univariate statistical analysis shows strong separability for PVH but limited significance for NPVH, our machine learning pipeline, tailored for high-dimensional feature integration, identifies that coupling features are crucial for both tasks. We achieve an AUC of 0.891 for PVH and 0.728 for NPVH, suggesting that while PVH is near-linearly separable, NPVH discrimination benefits from modeling non-linear feature interactions.

Audio effects (FX) shape sound in contemporary music practice. However, most interfaces present them as discrete modules and parameters that favor targeted adjustment over exploratory listening. This separation can make it difficult to build intuition about the broader space of possible transformations or to move fluidly between searching and refinement. We present FXplorer, an interface that organizes audio effects within a perceptually informed 2D space, allowing sound transformations to be browsed as a continuous landscape rather than as isolated presets. By combining established spatial interaction approaches and interpretable DAW-style controls with recent embedding-based machine learning methods for similarity and semantic search, the system brings exploration and parameter refinement into a single workspace. FXplorer supports composition, production, or performance by allowing users to edit and interpolate between effect presets interactively.

Acoustic metamaterial (AMM) inverse design is particularly challenging for broadband target responses due to acoustic dispersion: a structure that matches the desired response at one frequency may deviate at others, and modifying geometry to improve one sub-band often perturbs neighboring sub-bands. Yet existing broadband inverse-design approaches are either constrained by predefined templates, or rely on image representations that fail to preserve the geometric precision and structural connectivity required by acoustic structures. We present MetaSeq, a physics-guided, sequence-based generative framework for acoustic metamaterial inverse design. At its core, MetaSeq introduces a language that represents each AMM as a structured sequence, rather than as a pixel grid or fixed template. This representation preserves precise geometry, explicitly encodes connectivity, and casts inverse design as a sequence-to-sequence task from target response to structure sequence. MetaSeq further constructs a balanced, high-fidelity dataset with efficient calibration and complexity-based sampling. To address the one-to-many nature of inverse design, MetaSeq combines supervised pretraining with reinforcement learning fine-tuning guided by a physics-based solver and validity checker. Extensive evaluations against COMSOL and five baselines show that MetaSeq reduces response error by 45% over the best baseline.

Parkinson's disease (PD) is a progressive neurodegenerative disorder that frequently causes speech impairments associated with hypokinetic dysarthria. As speech production relies on the precise coordination of complex neuromuscular mechanisms, speech analysis has emerged as a promising non-invasive and cost-effective biomarker for early PD detection. Recent deep learning approaches have shown encouraging results; however, most existing methods rely on a single speech representation, potentially overlooking complementary pathological information encoded across different feature spaces. In this work, we propose a multi-branch deep learning framework for automatic PD detection from speech. Each recording is segmented into 5-second chunks and represented using three complementary modalities: Log-Mel spectrograms, MFCCs, and HuBERT embeddings extracted from raw waveforms. The spectrograms are processed using a pre-trained ResNet-18 encoder, MFCC sequences are modeled through a BiLSTM network, and raw speech is encoded using a pre-trained HuBERT model. To effectively integrate these heterogeneous representations, we introduce a context-guided cross-modal attention mechanism that dynamically weights temporal HuBERT embeddings according to the global acoustic context derived from the spectrogram and MFCC branches. Experiments conducted on the publicly available Spanish PC-GITA corpus under strict speaker-independent 5-fold cross-validation demonstrate the effectiveness of the proposed approach. The proposed architecture achieves an accuracy of 91.51%, an F1-score of 91.24%, and an AUC of 95.97%. Furthermore, ablation studies confirm the contribution of both the proposed context-guided cross-modal attention mechanism and the integration of complementary speech representations. These findings highlight the potential of heterogeneous speech modeling for robust and clinically reliable PD detection.

This study addresses the challenges composers and sound designers face in creating and refining tools to achieve their musical goals. Using evolutionary processes to promote diversity and foster serendipitous discoveries, we automate the search through uncharted sonic spaces for sound discovery, arguing that diversity-promoting algorithms can bridge the gap between the theoretical realisation and practical accessibility of sounds. We describe a system for generative sound synthesis combining Quality Diversity (QD) algorithms with a supervised discriminative model, inspired by the Innovation Engine algorithm, and explore different configurations and the interplay between the chosen synthesis approach and the discriminative model. We examine the interaction between Compositional Pattern Producing Networks (CPPNs) and Digital Signal Processing (DSP) graphs, introducing a novel approach that uses multiple specialised CPPNs for different frequency ranges; this yields simpler networks while maintaining performance comparable to single-CPPN setups. We also investigate evolutionary stepping stones by analysing goal switches between musical and non-musical contexts, revealing how lineages traverse unlikely paths to current elites. Expanding the behaviour space of a previous study to include various sound durations, we uncover specialisation within temporal niches. Results indicate that CPPN and DSP graphs coupled with a Multi-dimensional Archive of Phenotypic Elites (MAP-Elites) and a deep learning classifier can generate a substantial variety of synthetic sounds, diverse and innovative across temporal and contextual dimensions. We present the generated sound objects through an online explorer and as rendered sound files, and, in the context of music composition, an experimental application that showcases their creative potential across various durations and contexts.

Adaptive beamforming is a cornerstone of array signal processing, yet its performance often collapses in the face of complex, rapidly changing interference. When interferers appear or move unpredictably, conventional estimators encounter a fundamental memory trade-off: short windows enable rapid tracking but suffer from high estimation variance, while long windows provide stable rejection but fail to adapt to shifts. This challenge is resolved by introducing the Universal Switching Beamformer (USB), which integrates competitive sequential prediction into the beamforming architecture. By employing a linear transition diagram, the USB implicitly maintains an exponentially large family of candidate covariance histories and dynamically re-weights them based on their cumulative output power. This mechanism allows the beamformer to automatically vary its effective memory length without explicit change detection or heuristic parameter tuning. A theoretical upper bound is proven on the regret relative to an omniscient oracle that selects the best piecewise-stationary covariance model in hindsight. Extensive simulations and experiments on the SwellEx-96 dataset demonstrate that the USB achieves the agility of short-window estimators and the precision of long-term integration, providing a principled solution for tracking highly non-stationary scenes.

Recently, Transformers (e.g., Audio Spectrogram Transformers, AST) and state-space models (e.g., Audio Mamba, AuM) have achieved remarkable progress in audio modeling. However, the O(L^2) computational complexity of the Transformer architecture hinders efficient long-sequence processing, while the Mamba architecture tends to become unstable when scaling parameters and data. To address these challenges, this paper proposes AudioRWKV (A-RWKV), a highly efficient and stable architecture for audio modeling. Specifically, we inherit the stable and efficient recurrent formulation of RWKV7 and replace its 1D token-shift operation with a 2D depthwise separable convolution to better capture local spectro-temporal patterns. Furthermore, we adapt the original causal WKV kernel into a bidirectional WKV kernel (Bi-WKV), enabling global context modeling over the entire audio sequence while maintaining linear computational complexity. Benefiting from the inherent stability of the RWKV7 foundation, A-RWKV scales seamlessly to larger model sizes. Experimental results demonstrate that, under the same linear-model regime, A-RWKV-S (22M) achieves performance parity with AuM-B (92M) while exhibiting more stable throughput than AST; for long-form audio (~5 minutes 28 seconds), WKV7 achieves up to a 13.3X speedup in processing.

Sparse Autoencoders (SAEs) are powerful tools for interpreting neural representations, yet their use in audio remains underexplored. We train SAEs across all encoder layers of Whisper and HuBERT, provide an extensive evaluation of their stability, interpretability, and show their practical utility. Over 50% of the features remain consistent across random seeds, and reconstruction quality is preserved. SAE features capture general acoustic and semantic information as well as specific events, including environmental noises and paralinguistic sounds (e.g. laughter, whispering) and disentangle them effectively, requiring removal of only 19-27% of features to erase a concept. Feature steering reduces Whisper's false speech detections by 70% with negligible WER increase, demonstrating real-world applicability. Finally, we find SAE features correlated with human EEG activity during speech perception, indicating alignment with human neural processing. The code and checkpoints are available at https://github.com/audiosae/audiosae_demo.

The rapid advancement of speech generation models has heightened privacy and security concerns related to voice cloning (VC). Recent studies have investigated disrupting unauthorized voice cloning by introducing adversarial perturbations. However, determined attackers can mitigate these protective perturbations and successfully execute VC. In this study, we conduct the first systematic evaluation of these protective perturbations against VC under realistic threat models that include perturbation purification. Our findings reveal that while existing purification methods can neutralize a considerable portion of the protective perturbations, they still lead to distortions in the feature space of VC models, which degrades the performance of VC. From this perspective, we propose a novel two-stage purification method: (1) Purify the perturbed speech; (2) Refine it using phoneme guidance to align it with the clean speech distribution. Experimental results demonstrate that our method outperforms state-of-the-art purification methods in disrupting VC defenses. Our study reveals the limitations of adversarial perturbation-based VC defenses and underscores the urgent need for more robust solutions to mitigate the security and privacy risks posed by VC. The code and audio samples are available at https://de-antifake.github.io.