分享

Detecting and mitigating bias in speaker verification systems is important, as datasets, processing choices and algorithms can lead to performance differences that systematically favour some groups of people while disadvantaging others. Prior studies have thus measured performance differences across groups to evaluate bias. However, when comparing results across studies, it becomes apparent that they draw contradictory conclusions, hindering progress in this area. In this paper we investigate how measurement impacts the outcomes of bias evaluations. We show empirically that bias evaluations are strongly influenced by base metrics that measure performance, by the choice of ratio or difference-based bias measure, and by the aggregation of bias measures into meta-measures. Based on our findings, we recommend the use of ratio-based bias measures, in particular when the values of base metrics are small, or when base metrics with different orders of magnitude need to be compared.

Speech large language models (speech-LLMs) integrate speech and text-based foundation models to provide a unified framework for handling a wide range of downstream tasks. In this paper, we introduce WHISMA, a speech-LLM tailored for spoken language understanding (SLU) that demonstrates robust performance in various zero-shot settings. WHISMA combines the speech encoder from Whisper with the Llama-3 LLM, and is fine-tuned in a parameter-efficient manner on a comprehensive collection of SLU-related datasets. Our experiments show that WHISMA significantly improves the zero-shot slot filling performance on the SLURP benchmark, achieving a relative gain of 26.6% compared to the current state-of-the-art model. Furthermore, to evaluate WHISMA's generalisation capabilities to unseen domains, we develop a new task-agnostic benchmark named SLU-GLUE. The evaluation results indicate that WHISMA outperforms an existing speech-LLM (Qwen-Audio) with a relative gain of 33.0%.

We propose and analyze the use of an explicit time-context window for neural network-based spectral masking speech enhancement to leverage signal context dependencies between neighboring frames. In particular, we concentrate on soft masking and loss computed on the time-frequency representation of the reconstructed speech. We show that the application of a time-context windowing function at both input and output of the neural network model improves the soft mask estimation process by combining multiple estimates taken from different contexts. The proposed approach is only applied as post-optimization in inference mode, not requiring additional layers or special training for the neural network model. Our results show that the method consistently increases both intelligibility and signal quality of the denoised speech, as demonstrated for two classes of convolutional-based speech enhancement models. Importantly, the proposed method requires only a negligible ($\leq1\%$) increase in the number of model parameters, making it suitable for hardware-constrained applications.

Rap, a prominent genre of vocal performance, remains underexplored in vocal generation. General vocal synthesis depends on precise note and duration inputs, requiring users to have related musical knowledge, which limits flexibility. In contrast, rap typically features simpler melodies, with a core focus on a strong rhythmic sense that harmonizes with accompanying beats. In this paper, we propose Freestyler, the first system that generates rapping vocals directly from lyrics and accompaniment inputs. Freestyler utilizes language model-based token generation, followed by a conditional flow matching model to produce spectrograms and a neural vocoder to restore audio. It allows a 3-second prompt to enable zero-shot timbre control. Due to the scarcity of publicly available rap datasets, we also present RapBank, a rap song dataset collected from the internet, alongside a meticulously designed processing pipeline. Experimental results show that Freestyler produces high-quality rapping voice generation with enhanced naturalness and strong alignment with accompanying beats, both stylistically and rhythmically.

Aphasia is a language disorder affecting one third of stroke patients. Current aphasia assessment does not consider natural speech due to the time consuming nature of manual transcriptions and a lack of knowledge on how to analyze such data. Here, we evaluate the potential of automatic speech recognition (ASR) to transcribe Dutch aphasic speech and the ability of natural speech features to detect aphasia. A picture-description task was administered and automatically transcribed in 62 persons with aphasia and 57 controls. Acoustic and linguistic features were semi-automatically extracted and provided as input to a support vector machine (SVM) classifier. Our ASR model obtained a WER of 24.5%, outperforming earlier ASR models for aphasia. The SVM shows high accuracy (86.6%) at the individual level, with fluency features as most dominant to detect aphasia. ASR and semi-automatic feature extraction can thus facilitate natural speech analysis in a time efficient manner in clinical practice.

With the advancement of artificial intelligence (AI) technology, group-level emotion recognition (GER) has emerged as an important area in analyzing human behavior. Early GER methods are primarily relied on handcrafted features. However, with the proliferation of Deep Learning (DL) techniques and their remarkable success in diverse tasks, neural networks have garnered increasing interest in GER. Unlike individual's emotion, group emotions exhibit diversity and dynamics. Presently, several DL approaches have been proposed to effectively leverage the rich information inherent in group-level image and enhance GER performance significantly. In this survey, we present a comprehensive review of DL techniques applied to GER, proposing a new taxonomy for the field cover all aspects of GER based on DL. The survey overviews datasets, the deep GER pipeline, and performance comparisons of the state-of-the-art methods past decade. Moreover, it summarizes and discuss the fundamental approaches and advanced developments for each aspect. Furthermore, we identify outstanding challenges and suggest potential avenues for the design of robust GER systems. To the best of our knowledge, thus survey represents the first comprehensive review of deep GER methods, serving as a pivotal references for future GER research endeavors.

Fine-tuning large language models on private data for downstream applications poses significant privacy risks in potentially exposing sensitive information. Several popular community platforms now offer convenient distribution of a large variety of pre-trained models, allowing anyone to publish without rigorous verification. This scenario creates a privacy threat, as pre-trained models can be intentionally crafted to compromise the privacy of fine-tuning datasets. In this study, we introduce a novel poisoning technique that uses model-unlearning as an attack tool. This approach manipulates a pre-trained language model to increase the leakage of private data during the fine-tuning process. Our method enhances both membership inference and data extraction attacks while preserving model utility. Experimental results across different models, datasets, and fine-tuning setups demonstrate that our attacks significantly surpass baseline performance. This work serves as a cautionary note for users who download pre-trained models from unverified sources, highlighting the potential risks involved.

Automatic Speaker Verification (ASV) systems, which identify speakers based on their voice characteristics, have numerous applications, such as user authentication in financial transactions, exclusive access control in smart devices, and forensic fraud detection. However, the advancement of deep learning algorithms has enabled the generation of synthetic audio through Text-to-Speech (TTS) and Voice Conversion (VC) systems, exposing ASV systems to potential vulnerabilities. To counteract this, we propose a novel architecture named AASIST3. By enhancing the existing AASIST framework with Kolmogorov-Arnold networks, additional layers, encoders, and pre-emphasis techniques, AASIST3 achieves a more than twofold improvement in performance. It demonstrates minDCF results of 0.5357 in the closed condition and 0.1414 in the open condition, significantly enhancing the detection of synthetic voices and improving ASV security.

Currently, 6D pose estimation methods are benchmarked on datasets that consider, for their ground truth annotations, visual ambiguities as only related to global object symmetries. However, as previously observed [26], visual ambiguities can also happen depending on the viewpoint or the presence of occluding objects, when disambiguating parts become hidden. The visual ambiguities are therefore actually different across images. We thus first propose an automatic method to re-annotate those datasets with a 6D pose distribution specific to each image, taking into account the visibility of the object surface in the image to correctly determine the visual ambiguities. Given this improved ground truth, we re-evaluate the state-of-the-art methods and show this greatly modify the ranking of these methods. Our annotations also allow us to benchmark recent methods able to estimate a pose distribution on real images for the first time. We will make our annotations for the T-LESS dataset and our code publicly available.

Knowledge editing has been proposed as an effective method for updating and correcting the internal knowledge of Large Language Models (LLMs). However, existing editing methods often struggle with complex tasks, such as multi-hop reasoning. In this paper, we identify and investigate the phenomenon of Editing Overfit, where edited models assign disproportionately high probabilities to the edit target, hindering the generalization of new knowledge in complex scenarios. We attribute this issue to the current editing paradigm, which places excessive emphasis on the direct correspondence between the input prompt and the edit target for each edit sample. To further explore this issue, we introduce a new benchmark, EVOKE (EValuation of Editing Overfit in Knowledge Editing), along with fine-grained evaluation metrics. Through comprehensive experiments and analysis, we demonstrate that Editing Overfit is prevalent in current editing methods and that common overfitting mitigation strategies are of limited effectiveness in knowledge editing. To overcome this, inspired by LLMs' knowledge recall mechanisms, we propose a new plug-and-play strategy called Learn to Inference (LTI), which introduce a Multi-stage Inference Constraint module to guide the edited models in recalling new knowledge similarly to how unedited LLMs leverage knowledge through in-context learning. Extensive experimental results across a wide range of tasks validate the effectiveness of LTI in mitigating Editing Overfit.

The rapid proliferation of AI-manipulated or generated audio deepfakes poses serious challenges to media integrity and election security. Current AI-driven detection solutions lack explainability and underperform in real-world settings. In this paper, we introduce novel explainability methods for state-of-the-art transformer-based audio deepfake detectors and open-source a novel benchmark for real-world generalizability. By narrowing the explainability gap between transformer-based audio deepfake detectors and traditional methods, our results not only build trust with human experts, but also pave the way for unlocking the potential of citizen intelligence to overcome the scalability issue in audio deepfake detection.

Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model--a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining.