With the advancement of generative modeling techniques, synthetic human speech becomes increasingly indistinguishable from real, and tricky challenges are elicited for the audio deepfake detection (ADD) system. In this paper, we exploit audio features to improve the generalizability of ADD systems. Investigation of the ADD task performance is conducted over a broad range of audio features, including various handcrafted features and learning-based features. Experiments show that learning-based audio features pretrained on a large amount of data generalize better than hand-crafted features on out-of-domain scenarios. Subsequently, we further improve the generalizability of the ADD system using proposed multi-feature approaches to incorporate complimentary information from features of different views. The model trained on ASV2019 data achieves an equal error rate of 24.27\% on the In-the-Wild dataset.
Emotional Voice Messages (EMOVOME) is a spontaneous speech dataset containing 999 audio messages from real conversations on a messaging app from 100 Spanish speakers, gender balanced. Voice messages were produced in-the-wild conditions before participants were recruited, avoiding any conscious bias due to laboratory environment. Audios were labeled in valence and arousal dimensions by three non-experts and two experts, which were then combined to obtain a final label per dimension. The experts also provided an extra label corresponding to seven emotion categories. To set a baseline for future investigations using EMOVOME, we implemented emotion recognition models using both speech and audio transcriptions. For speech, we used the standard eGeMAPS feature set and support vector machines, obtaining 49.27% and 44.71% unweighted accuracy for valence and arousal respectively. For text, we fine-tuned a multilingual BERT model and achieved 61.15% and 47.43% unweighted accuracy for valence and arousal respectively. This database will significantly contribute to research on emotion recognition in the wild, while also providing a unique natural and freely accessible resource for Spanish.
Probing the memorization of large language models holds significant importance. Previous works have established metrics for quantifying memorization, explored various influencing factors, such as data duplication, model size, and prompt length, and evaluated memorization by comparing model outputs with training corpora. However, the training corpora are of enormous scale and its pre-processing is time-consuming. To explore memorization without accessing training data, we propose a novel approach, named ROME, wherein memorization is explored by comparing disparities across memorized and non-memorized. Specifically, models firstly categorize the selected samples into memorized and non-memorized groups, and then comparing the demonstrations in the two groups from the insights of text, probability, and hidden state. Experimental findings show the disparities in factors including word length, part-of-speech, word frequency, mean and variance, just to name a few.
Stance detection of social media text is a key component of downstream tasks involving the identification of groups of users with opposing opinions on contested topics such as vaccination and within arguments. In particular, stance provides an indication of an opinion towards an entity. This paper introduces DIVERSE, a dataset of over 173,000 YouTube video comments annotated for their stance towards videos of the U.S. military. The stance is annotated through a human-guided, machine-assisted labeling methodology that makes use of weak signals of tone within the sentence as supporting indicators, as opposed to using manual annotations by humans. These weak signals consist of the presence of hate speech and sarcasm, the presence of specific keywords, the sentiment of the text, and the stance inference from two Large Language Models. The weak signals are then consolidated using a data programming model before each comment is annotated with a final stance label. On average, the videos have 200 comments each, and the stance of the comments skews slightly towards the "against" characterization for both the U.S. Army and the videos posted on the channel.
In recent years, neural network-based Wake Word Spotting achieves good performance on clean audio samples but struggles in noisy environments. Audio-Visual Wake Word Spotting (AVWWS) receives lots of attention because visual lip movement information is not affected by complex acoustic scenes. Previous works usually use simple addition or concatenation for multi-modal fusion. The inter-modal correlation remains relatively under-explored. In this paper, we propose a novel module called Frame-Level Cross-Modal Attention (FLCMA) to improve the performance of AVWWS systems. This module can help model multi-modal information at the frame-level through synchronous lip movements and speech signals. We train the end-to-end FLCMA based Audio-Visual Conformer and further improve the performance by fine-tuning pre-trained uni-modal models for the AVWWS task. The proposed system achieves a new state-of-the-art result (4.57% WWS score) on the far-field MISP dataset.
While considerable progress has been made in achieving accurate lip synchronization for 3D speech-driven talking face generation, the task of incorporating expressive facial detail synthesis aligned with the speaker's speaking status remains challenging. Our goal is to directly leverage the inherent style information conveyed by human speech for generating an expressive talking face that aligns with the speaking status. In this paper, we propose AVI-Talking, an Audio-Visual Instruction system for expressive Talking face generation. This system harnesses the robust contextual reasoning and hallucination capability offered by Large Language Models (LLMs) to instruct the realistic synthesis of 3D talking faces. Instead of directly learning facial movements from human speech, our two-stage strategy involves the LLMs first comprehending audio information and generating instructions implying expressive facial details seamlessly corresponding to the speech. Subsequently, a diffusion-based generative network executes these instructions. This two-stage process, coupled with the incorporation of LLMs, enhances model interpretability and provides users with flexibility to comprehend instructions and specify desired operations or modifications. Extensive experiments showcase the effectiveness of our approach in producing vivid talking faces with expressive facial movements and consistent emotional status.
Brain-computer interface (BCI) technology facilitates communication between the human brain and computers, primarily utilizing electroencephalography (EEG) signals to discern human intentions. Although EEG-based BCI systems have been developed for paralysis individuals, ongoing studies explore systems for speech imagery and motor imagery (MI). This study introduces FingerNet, a specialized network for fine MI classification, departing from conventional gross MI studies. The proposed FingerNet could extract spatial and temporal features from EEG signals, improving classification accuracy within the same hand. The experimental results demonstrated that performance showed significantly higher accuracy in classifying five finger-tapping tasks, encompassing thumb, index, middle, ring, and little finger movements. FingerNet demonstrated dominant performance compared to the conventional baseline models, EEGNet and DeepConvNet. The average accuracy for FingerNet was 0.3049, whereas EEGNet and DeepConvNet exhibited lower accuracies of 0.2196 and 0.2533, respectively. Statistical validation also demonstrates the predominance of FingerNet over baseline networks. For biased predictions, particularly for thumb and index classes, we led to the implementation of weighted cross-entropy and also adapted the weighted cross-entropy, a method conventionally employed to mitigate class imbalance. The proposed FingerNet involves optimizing network structure, improving performance, and exploring applications beyond fine MI. Moreover, the weighted Cross Entropy approach employed to address such biased predictions appears to have broader applicability and relevance across various domains involving multi-class classification tasks. We believe that effective execution of motor imagery can be achieved not only for fine MI, but also for local muscle MI
Self-supervised learning (SSL) for automated speech recognition in terms of its emotional content, can be heavily degraded by the presence noise, affecting the efficiency of modeling the intricate temporal and spectral informative structures of speech. Recently, SSL on large speech datasets, as well as new audio-specific SSL proxy tasks, such as, temporal and frequency masking, have emerged, yielding superior performance compared to classic approaches drawn from the image augmentation domain. Our proposed contribution builds upon this successful paradigm by introducing CochCeps-Augment, a novel bio-inspired masking augmentation task for self-supervised contrastive learning of speech representations. Specifically, we utilize the newly introduced bio-inspired cochlear cepstrogram (CCGRAM) to derive noise robust representations of input speech, that are then further refined through a self-supervised learning scheme. The latter employs SimCLR to generate contrastive views of a CCGRAM through masking of its angle and quefrency dimensions. Our experimental approach and validations on the emotion recognition K-EmoCon benchmark dataset, for the first time via a speaker-independent approach, features unsupervised pre-training, linear probing and fine-tuning. Our results potentiate CochCeps-Augment to serve as a standard tool in speech emotion recognition analysis, showing the added value of incorporating bio-inspired masking as an informative augmentation task for self-supervision. Our code for implementing CochCeps-Augment will be made available at: https://github.com/GiannisZgs/CochCepsAugment.
The accuracy of modern automatic speaker verification (ASV) systems, when trained exclusively on adult data, drops substantially when applied to children's speech. The scarcity of children's speech corpora hinders fine-tuning ASV systems for children's speech. Hence, there is a timely need to explore more effective ways of reusing adults' speech data. One promising approach is to align vocal-tract parameters between adults and children through children-specific data augmentation, referred here to as ChildAugment. Specifically, we modify the formant frequencies and formant bandwidths of adult speech to emulate children's speech. The modified spectra are used to train ECAPA-TDNN (emphasized channel attention, propagation, and aggregation in time-delay neural network) recognizer for children. We compare ChildAugment against various state-of-the-art data augmentation techniques for children's ASV. We also extensively compare different scoring methods, including cosine scoring, PLDA (probabilistic linear discriminant analysis), and NPLDA (neural PLDA). We also propose a low-complexity weighted cosine score for extremely low-resource children ASV. Our findings on the CSLU kids corpus indicate that ChildAugment holds promise as a simple, acoustics-motivated approach, for improving state-of-the-art deep learning based ASV for children. We achieve up to 12.45% (boys) and 11.96% (girls) relative improvement over the baseline.
The automatic identification of offensive language such as hate speech is important to keep discussions civil in online communities. Identifying hate speech in multimodal content is a particularly challenging task because offensiveness can be manifested in either words or images or a juxtaposition of the two. This paper presents the MasonPerplexity submission for the Shared Task on Multimodal Hate Speech Event Detection at CASE 2024 at EACL 2024. The task is divided into two sub-tasks: sub-task A focuses on the identification of hate speech and sub-task B focuses on the identification of targets in text-embedded images during political events. We use an XLM-roBERTa-large model for sub-task A and an ensemble approach combining XLM-roBERTa-base, BERTweet-large, and BERT-base for sub-task B. Our approach obtained 0.8347 F1-score in sub-task A and 0.6741 F1-score in sub-task B ranking 3rd on both sub-tasks.