From Sharpness to Better Generalization for Speech Deepfake Detection

Generalization remains a critical challenge in speech deepfake detection (SDD). While various approaches aim to improve robustness, generalization is typically assessed through performance metrics like equal error rate without a theoretical framework to explain model performance. This work investigates sharpness as a theoretical proxy for generalization in SDD. We analyze how sharpness responds to domain shifts and find it increases in unseen conditions, indicating higher model sensitivity. Based on this, we apply Sharpness-Aware Minimization (SAM) to reduce sharpness explicitly, leading to better and more stable performance across diverse unseen test sets. Furthermore, correlation analysis confirms a statistically significant relationship between sharpness and generalization in most test settings. These findings suggest that sharpness can serve as a theoretical indicator for generalization in SDD and that sharpness-aware training offers a promising strategy for improving robustness.

Improving Speech Enhancement with Multi-Metric Supervision from Learned Quality Assessment

Speech quality assessment (SQA) aims to predict the perceived quality of speech signals under a wide range of distortions. It is inherently connected to speech enhancement (SE), which seeks to improve speech quality by removing unwanted signal components. While SQA models are widely used to evaluate SE performance, their potential to guide SE training remains underexplored. In this work, we investigate a training framework that leverages a SQA model, trained to predict multiple evaluation metrics from a public SE leaderboard, as a supervisory signal for SE. This approach addresses a key limitation of conventional SE objectives, such as SI-SNR, which often fail to align with perceptual quality and generalize poorly across evaluation metrics. Moreover, it enables training on real-world data where clean references are unavailable. Experiments on both simulated and real-world test sets show that SQA-guided training consistently improves performance across a range of quality metrics.

DenoiseRotator: Enhance Pruning Robustness for LLMs via Importance Concentration

Pruning is a widely used technique to compress large language models (LLMs) by removing unimportant weights, but it often suffers from significant performance degradation - especially under semi-structured sparsity constraints. Existing pruning methods primarily focus on estimating the importance of individual weights, which limits their ability to preserve critical capabilities of the model. In this work, we propose a new perspective: rather than merely selecting which weights to prune, we first redistribute parameter importance to make the model inherently more amenable to pruning. By minimizing the information entropy of normalized importance scores, our approach concentrates importance onto a smaller subset of weights, thereby enhancing pruning robustness. We instantiate this idea through DenoiseRotator, which applies learnable orthogonal transformations to the model's weight matrices. Our method is model-agnostic and can be seamlessly integrated with existing pruning techniques such as Magnitude, SparseGPT, and Wanda. Evaluated on LLaMA3, Qwen2.5, and Mistral models under 50% unstructured and 2:4 semi-structured sparsity, DenoiseRotator consistently improves perplexity and zero-shot accuracy. For instance, on LLaMA3-70B pruned with SparseGPT at 2:4 semi-structured sparsity, DenoiseRotator reduces the perplexity gap to the dense model by 58%, narrowing the degradation from 8.1 to 3.4 points. Codes are available at https://github.com/Axel-gu/DenoiseRotator.

Zero-Shot Streaming Text to Speech Synthesis with Transducer and Auto-Regressive Modeling

Zero-shot streaming text-to-speech is an important research topic in human-computer interaction. Existing methods primarily use a lookahead mechanism, relying on future text to achieve natural streaming speech synthesis, which introduces high processing latency. To address this issue, we propose SMLLE, a streaming framework for generating high-quality speech frame-by-frame. SMLLE employs a Transducer to convert text into semantic tokens in real time while simultaneously obtaining duration alignment information. The combined outputs are then fed into a fully autoregressive (AR) streaming model to reconstruct mel-spectrograms. To further stabilize the generation process, we design a Delete < Bos > Mechanism that allows the AR model to access future text introducing as minimal delay as possible. Experimental results suggest that the SMLLE outperforms current streaming TTS methods and achieves comparable performance over sentence-level TTS systems. Samples are available on https://anonymous.4open.science/w/demo_page-48B7/.

BR-ASR: Efficient and Scalable Bias Retrieval Framework for Contextual Biasing ASR in Speech LLM

While speech large language models (SpeechLLMs) have advanced standard automatic speech recognition (ASR), contextual biasing for named entities and rare words remains challenging, especially at scale. To address this, we propose BR-ASR: a Bias Retrieval framework for large-scale contextual biasing (up to 200k entries) via two innovations: (1) speech-and-bias contrastive learning to retrieve semantically relevant candidates; (2) dynamic curriculum learning that mitigates homophone confusion which negatively impacts the final performance. The is a general framework that allows seamless integration of the retrieved candidates into diverse ASR systems without fine-tuning. Experiments on LibriSpeech test-clean/-other achieve state-of-the-art (SOTA) biased word error rates (B-WER) of 2.8%/7.1% with 2000 bias words, delivering 45% relative improvement over prior methods. BR-ASR also demonstrates high scalability: when expanding the bias list to 200k where traditional methods generally fail, it induces only 0.3 / 2.9% absolute WER / B-WER degradation with a 99.99% pruning rate and only 20ms latency per query on test-other.

Advanced Zero-Shot Text-to-Speech for Background Removal and Preservation with Controllable Masked Speech Prediction

The acoustic background plays a crucial role in natural conversation. It provides context and helps listeners understand the environment, but a strong background makes it difficult for listeners to understand spoken words. The appropriate handling of these backgrounds is situation-dependent: Although it may be necessary to remove background to ensure speech clarity, preserving the background is sometimes crucial to maintaining the contextual integrity of the speech. Despite recent advancements in zero-shot Text-to-Speech technologies, current systems often struggle with speech prompts containing backgrounds. To address these challenges, we propose a Controllable Masked Speech Prediction strategy coupled with a dual-speaker encoder, utilizing a task-related control signal to guide the prediction of dual background removal and preservation targets. Experimental results demonstrate that our approach enables precise control over the removal or preservation of background across various acoustic conditions and exhibits strong generalization capabilities in unseen scenarios.

Generalizable Audio Deepfake Detection via Latent Space Refinement and Augmentation

Advances in speech synthesis technologies, like text-to-speech (TTS) and voice conversion (VC), have made detecting deepfake speech increasingly challenging. Spoofing countermeasures often struggle to generalize effectively, particularly when faced with unseen attacks. To address this, we propose a novel strategy that integrates Latent Space Refinement (LSR) and Latent Space Augmentation (LSA) to improve the generalization of deepfake detection systems. LSR introduces multiple learnable prototypes for the spoof class, refining the latent space to better capture the intricate variations within spoofed data. LSA further diversifies spoofed data representations by applying augmentation techniques directly in the latent space, enabling the model to learn a broader range of spoofing patterns. We evaluated our approach on four representative datasets, i.e. ASVspoof 2019 LA, ASVspoof 2021 LA and DF, and In-The-Wild. The results show that LSR and LSA perform well individually, and their integration achieves competitive results, matching or surpassing current state-of-the-art methods.

SLIDE: Integrating Speech Language Model with LLM for Spontaneous Spoken Dialogue Generation

Recently, ``textless" speech language models (SLMs) based on speech units have made huge progress in generating naturalistic speech, including non-verbal vocalizations. However, the generated speech samples often lack semantic coherence. In this paper, we propose SLM and LLM Integration for spontaneous spoken Dialogue gEneration (SLIDE). Specifically, we first utilize an LLM to generate the textual content of spoken dialogue. Next, we convert the textual dialogues into phoneme sequences and use a two-tower transformer-based duration predictor to predict the duration of each phoneme. Finally, an SLM conditioned on the spoken phoneme sequences is used to vocalize the textual dialogue. Experimental results on the Fisher dataset demonstrate that our system can generate naturalistic spoken dialogue while maintaining high semantic coherence.

Scale This, Not That: Investigating Key Dataset Attributes for Efficient Speech Enhancement Scaling

Recent speech enhancement models have shown impressive performance gains by scaling up model complexity and training data. However, the impact of dataset variability (e.g. text, language, speaker, and noise) has been underexplored. Analyzing each attribute individually is often challenging, as multiple attributes are usually entangled in commonly used datasets, posing a significant obstacle in understanding the distinct contributions of each attribute to the model's performance. To address this challenge, we propose a generation-training-evaluation framework that leverages zero-shot text-to-speech systems to investigate the impact of controlled attribute variations on speech enhancement performance. It enables us to synthesize training datasets in a scalable manner while carefully altering each attribute. Based on the proposed framework, we analyze the scaling effects of various dataset attributes on the performance of both discriminative and generative SE models. Extensive experiments on multi-domain corpora imply that acoustic attributes (e.g., speaker and noise) are much more important to current speech enhancement models than semantic attributes (e.g., language and text), offering new insights for future research.

Memory-Efficient Training for Deep Speaker Embedding Learning in Speaker Verification

Recent speaker verification (SV) systems have shown a trend toward adopting deeper speaker embedding extractors. Although deeper and larger neural networks can significantly improve performance, their substantial memory requirements hinder training on consumer GPUs. In this paper, we explore a memory-efficient training strategy for deep speaker embedding learning in resource-constrained scenarios. Firstly, we conduct a systematic analysis of GPU memory allocation during SV system training. Empirical observations show that activations and optimizer states are the main sources of memory consumption. For activations, we design two types of reversible neural networks which eliminate the need to store intermediate activations during back-propagation, thereby significantly reducing memory usage without performance loss. For optimizer states, we introduce a dynamic quantization approach that replaces the original 32-bit floating-point values with a dynamic tree-based 8-bit data type. Experimental results on VoxCeleb demonstrate that the reversible variants of ResNets and DF-ResNets can perform training without the need to cache activations in GPU memory. In addition, the 8-bit versions of SGD and Adam save 75% of memory costs while maintaining performance compared to their 32-bit counterparts. Finally, a detailed comparison of memory usage and performance indicates that our proposed models achieve up to 16.2x memory savings, with nearly identical parameters and performance compared to the vanilla systems. In contrast to the previous need for multiple high-end GPUs such as the A100, we can effectively train deep speaker embedding extractors with just one or two consumer-level 2080Ti GPUs.