Deep Speech Synthesis from Multimodal Articulatory Representations

The amount of articulatory data available for training deep learning models is much less compared to acoustic speech data. In order to improve articulatory-to-acoustic synthesis performance in these low-resource settings, we propose a multimodal pre-training framework. On single-speaker speech synthesis tasks from real-time magnetic resonance imaging and surface electromyography inputs, the intelligibility of synthesized outputs improves noticeably. For example, compared to prior work, utilizing our proposed transfer learning methods improves the MRI-to-speech performance by 36% word error rate. In addition to these intelligibility results, our multimodal pre-trained models consistently outperform unimodal baselines on three objective and subjective synthesis quality metrics.

Investigating Effective Speaker Property Privacy Protection in Federated Learning for Speech Emotion Recognition

Federated Learning (FL) is a privacy-preserving approach that allows servers to aggregate distributed models transmitted from local clients rather than training on user data. More recently, FL has been applied to Speech Emotion Recognition (SER) for secure human-computer interaction applications. Recent research has found that FL is still vulnerable to inference attacks. To this end, this paper focuses on investigating the security of FL for SER concerning property inference attacks. We propose a novel method to protect the property information in speech data by decomposing various properties in the sound and adding perturbations to these properties. Our experiments show that the proposed method offers better privacy-utility trade-offs than existing methods. The trade-offs enable more effective attack prevention while maintaining similar FL utility levels. This work can guide future work on privacy protection methods in speech processing.

Speech as a Biomarker for Disease Detection

Speech is a rich biomarker that encodes substantial information about the health of a speaker, and thus it has been proposed for the detection of numerous diseases, achieving promising results. However, questions remain about what the models trained for the automatic detection of these diseases are actually learning and the basis for their predictions, which can significantly impact patients' lives. This work advocates for an interpretable health model, suitable for detecting several diseases, motivated by the observation that speech-affecting disorders often have overlapping effects on speech signals. A framework is presented that first defines "reference speech" and then leverages this definition for disease detection. Reference speech is characterized through reference intervals, i.e., the typical values of clinically meaningful acoustic and linguistic features derived from a reference population. This novel approach in the field of speech as a biomarker is inspired by the use of reference intervals in clinical laboratory science. Deviations of new speakers from this reference model are quantified and used as input to detect Alzheimer's and Parkinson's disease. The classification strategy explored is based on Neural Additive Models, a type of glass-box neural network, which enables interpretability. The proposed framework for reference speech characterization and disease detection is designed to support the medical community by providing clinically meaningful explanations that can serve as a valuable second opinion.

NeuroSpex: Neuro-Guided Speaker Extraction with Cross-Modal Attention

In the study of auditory attention, it has been revealed that there exists a robust correlation between attended speech and elicited neural responses, measurable through electroencephalography (EEG). Therefore, it is possible to use the attention information available within EEG signals to guide the extraction of the target speaker in a cocktail party computationally. In this paper, we present a neuro-guided speaker extraction model, i.e. NeuroSpex, using the EEG response of the listener as the sole auxiliary reference cue to extract attended speech from monaural speech mixtures. We propose a novel EEG signal encoder that captures the attention information. Additionally, we propose a cross-attention (CA) mechanism to enhance the speech feature representations, generating a speaker extraction mask. Experimental results on a publicly available dataset demonstrate that our proposed model outperforms two baseline models across various evaluation metrics.

On the Role of Visual Grounding in VQA

Visual Grounding (VG) in VQA refers to a model's proclivity to infer answers based on question-relevant image regions. Conceptually, VG identifies as an axiomatic requirement of the VQA task. In practice, however, DNN-based VQA models are notorious for bypassing VG by way of shortcut (SC) learning without suffering obvious performance losses in standard benchmarks. To uncover the impact of SC learning, Out-of-Distribution (OOD) tests have been proposed that expose a lack of VG with low accuracy. These tests have since been at the center of VG research and served as basis for various investigations into VG's impact on accuracy. However, the role of VG in VQA still remains not fully understood and has not yet been properly formalized. In this work, we seek to clarify VG's role in VQA by formalizing it on a conceptual level. We propose a novel theoretical framework called "Visually Grounded Reasoning" (VGR) that uses the concepts of VG and Reasoning to describe VQA inference in ideal OOD testing. By consolidating fundamental insights into VG's role in VQA, VGR helps to reveal rampant VG-related SC exploitation in OOD testing, which explains why the relationship between VG and OOD accuracy has been difficult to define. Finally, we propose an approach to create OOD tests that properly emphasize a requirement for VG, and show how to improve performance on them.

Speech Emotion Recognition under Resource Constraints with Data Distillation

Speech emotion recognition (SER) plays a crucial role in human-computer interaction. The emergence of edge devices in the Internet of Things (IoT) presents challenges in constructing intricate deep learning models due to constraints in memory and computational resources. Moreover, emotional speech data often contains private information, raising concerns about privacy leakage during the deployment of SER models. To address these challenges, we propose a data distillation framework to facilitate efficient development of SER models in IoT applications using a synthesised, smaller, and distilled dataset. Our experiments demonstrate that the distilled dataset can be effectively utilised to train SER models with fixed initialisation, achieving performances comparable to those developed using the original full emotional speech dataset.

Diff-ETS: Learning a Diffusion Probabilistic Model for Electromyography-to-Speech Conversion

Electromyography-to-Speech (ETS) conversion has demonstrated its potential for silent speech interfaces by generating audible speech from Electromyography (EMG) signals during silent articulations. ETS models usually consist of an EMG encoder which converts EMG signals to acoustic speech features, and a vocoder which then synthesises the speech signals. Due to an inadequate amount of available data and noisy signals, the synthesised speech often exhibits a low level of naturalness. In this work, we propose Diff-ETS, an ETS model which uses a score-based diffusion probabilistic model to enhance the naturalness of synthesised speech. The diffusion model is applied to improve the quality of the acoustic features predicted by an EMG encoder. In our experiments, we evaluated fine-tuning the diffusion model on predictions of a pre-trained EMG encoder, and training both models in an end-to-end fashion. We compared Diff-ETS with a baseline ETS model without diffusion using objective metrics and a listening test. The results indicated the proposed Diff-ETS significantly improved speech naturalness over the baseline.

STAA-Net: A Sparse and Transferable Adversarial Attack for Speech Emotion Recognition

Speech contains rich information on the emotions of humans, and Speech Emotion Recognition (SER) has been an important topic in the area of human-computer interaction. The robustness of SER models is crucial, particularly in privacy-sensitive and reliability-demanding domains like private healthcare. Recently, the vulnerability of deep neural networks in the audio domain to adversarial attacks has become a popular area of research. However, prior works on adversarial attacks in the audio domain primarily rely on iterative gradient-based techniques, which are time-consuming and prone to overfitting the specific threat model. Furthermore, the exploration of sparse perturbations, which have the potential for better stealthiness, remains limited in the audio domain. To address these challenges, we propose a generator-based attack method to generate sparse and transferable adversarial examples to deceive SER models in an end-to-end and efficient manner. We evaluate our method on two widely-used SER datasets, Database of Elicited Mood in Speech (DEMoS) and Interactive Emotional dyadic MOtion CAPture (IEMOCAP), and demonstrate its ability to generate successful sparse adversarial examples in an efficient manner. Moreover, our generated adversarial examples exhibit model-agnostic transferability, enabling effective adversarial attacks on advanced victim models.

Uncovering the Full Potential of Visual Grounding Methods in VQA

Visual Grounding (VG) methods in Visual Question Answering (VQA) attempt to improve VQA performance by strengthening a model's reliance on question-relevant visual information. The presence of such relevant information in the visual input is typically assumed in training and testing. This assumption, however, is inherently flawed when dealing with imperfect image representations common in large-scale VQA, where the information carried by visual features frequently deviates from expected ground-truth contents. As a result, training and testing of VG-methods is performed with largely inaccurate data, which obstructs proper assessment of their potential benefits. In this work, we demonstrate that current evaluation schemes for VG-methods are problematic due to the flawed assumption of availability of relevant visual information. Our experiments show that the potential benefits of these methods are severely underestimated as a result.

NeuroHeed: Neuro-Steered Speaker Extraction using EEG Signals

Humans possess the remarkable ability to selectively attend to a single speaker amidst competing voices and background noise, known as selective auditory attention. Recent studies in auditory neuroscience indicate a strong correlation between the attended speech signal and the corresponding brain's elicited neuronal activities, which the latter can be measured using affordable and non-intrusive electroencephalography (EEG) devices. In this study, we present NeuroHeed, a speaker extraction model that leverages EEG signals to establish a neuronal attractor which is temporally associated with the speech stimulus, facilitating the extraction of the attended speech signal in a cocktail party scenario. We propose both an offline and an online NeuroHeed, with the latter designed for real-time inference. In the online NeuroHeed, we additionally propose an autoregressive speaker encoder, which accumulates past extracted speech signals for self-enrollment of the attended speaker information into an auditory attractor, that retains the attentional momentum over time. Online NeuroHeed extracts the current window of the speech signals with guidance from both attractors. Experimental results demonstrate that NeuroHeed effectively extracts brain-attended speech signals, achieving high signal quality, excellent perceptual quality, and intelligibility in a two-speaker scenario.