Themos Stafylakis

Alexandros Xenos, Themos Stafylakis, Ioannis Patras et al.

This paper is on the problem of Knowledge-Based Visual Question Answering (KB-VQA). Recent works have emphasized the significance of incorporating both explicit (through external databases) and implicit (through LLMs) knowledge to answer questions requiring external knowledge effectively. A common limitation of such approaches is that they consist of relatively complicated pipelines and often heavily rely on accessing GPT-3 API. Our main contribution in this paper is to propose a much simpler and readily reproducible pipeline which, in a nutshell, is based on efficient in-context learning by prompting LLaMA (1 and 2) using question-informative captions as contextual information. Contrary to recent approaches, our method is training-free, does not require access to external databases or APIs, and yet achieves state-of-the-art accuracy on the OK-VQA and A-OK-VQA datasets. Finally, we perform several ablation studies to understand important aspects of our method. Our code is publicly available at https://github.com/alexandrosXe/ASimple-Baseline-For-Knowledge-Based-VQA

Junyi Peng, Oldřich Plchot, Xiao Song et al.

Modern speaker verification (SV) systems rely on speaker embeddings that are effective but difficult to interpret or query in natural language. Most existing speech-text corpora target controllable synthesis or utterance-level captioning, and provide limited speaker-level supervision for in-the-wild speaker recognition. This paper introduces SpeakerCard-1M, a bilingual speaker-centric resource for evidence-grounded SV, derived from VoxCeleb1/2 and CN-Celeb1/2, where the "-1M" suffix refers to the 1.78M utterance-level captions contained in the release. We adopt a tool-first, LLM-last approach: ten acoustic probes produce field-level evidence, the evidence is aggregated into speaker profiles under a schema that separates relatively stable traits from utterance-level states, and bilingual Speaker Cards are rendered by a constrained LLM that sees only the structured fields. The release includes 56.7K Speaker Card records over 10.2K speakers, 1.78M utterance-level captions, and speaker-ID-disjoint hard-negative triplets. We further define two SV-oriented cross-modal protocols, bidirectional Speaker-Text Retrieval (T2S-R / S2T-R) and Attribute-Conditioned Verification (AC-Verify), and compare a dual-encoder baseline against recent audio language models under a zero-shot forced-choice setting. Joint audio-text training increases VoxCeleb1-O EER by 0.31% absolute over the audio-only baseline. Under a style-symmetric LLM-generated counterfactual protocol, eight recent audio language models (7B-30B+ parameters, both open- and closed-source) score 49-77% on pitch-level AC-Verify under two-way forced choice, compared with 88.66% reached by our dual encoder.

Ioannis Prokopiou, Ioannis Sina, Agisilaos Kounelis et al.

The advancement of Machine learning (ML), Large Audio Language Models (LALMs), and autonomous AI agents in Music Information Retrieval (MIR) necessitates a shift from static tagging to rich, human-aligned representation learning. However, the scarcity of open-source infrastructure capable of capturing the subjective nuances of audio annotation remains a critical bottleneck. This paper introduces \textbf{LabelBuddy}, an open-source collaborative auto-tagging audio annotation tool designed to bridge the gap between human intent and machine understanding. Unlike static tools, it decouples the interface from inference via containerized backends, allowing users to plug in custom models for AI-assisted pre-annotation. We describe the system architecture, which supports multi-user consensus, containerized model isolation, and a roadmap for extending agents and LALMs. Code available at https://github.com/GiannisProkopiou/gsoc2022-Label-buddy.

Ioannis Prokopiou, Pantelis Vikatos, Maximos Kaliakatsos-Papakostas et al.

Transformer-based architectures have significantly advanced the generation of complex symbolic sequences, yet a significant gap remains in achieving fine-grained, interpretable control over discrete signal attributes. This paper investigates the mechanistic interpretability of the Multitrack Music Transformer (MMT) and proposes a framework for deterministic attribute modulation without retraining to bridge this gap via inference-time activation steering. Utilizing the Difference-in-Means (DiffMean) methodology, we isolate latent directions for signal attributes, specifically Pitch and Duration, within the residual stream. We validate the Linear Representation Hypothesis in this domain, achieving high correlation between steering magnitude and attribute shift. To address the inherent feature entanglement in multi-attribute steering, we introduce a Dual Steering framework utilizing Gram-Schmidt Orthogonalization. Experimental results demonstrate that this geometric decoupling reduces conceptual interference and signal degradation compared to naive vector addition, enabling independent deterministic control even against strong autoregressive conditioning.

Gaëlle Laperrière, Valentin Pelloin, Mickaël Rouvier et al.

In this paper we examine the use of semantically-aligned speech representations for end-to-end spoken language understanding (SLU). We employ the recently-introduced SAMU-XLSR model, which is designed to generate a single embedding that captures the semantics at the utterance level, semantically aligned across different languages. This model combines the acoustic frame-level speech representation learning model (XLS-R) with the Language Agnostic BERT Sentence Embedding (LaBSE) model. We show that the use of the SAMU-XLSR model instead of the initial XLS-R model improves significantly the performance in the framework of end-to-end SLU. Finally, we present the benefits of using this model towards language portability in SLU.

Niko Brümmer, Albert Swart, Ladislav Mošner et al.

In speaker recognition, where speech segments are mapped to embeddings on the unit hypersphere, two scoring backends are commonly used, namely cosine scoring or PLDA. Both have advantages and disadvantages, depending on the context. Cosine scoring follows naturally from the spherical geometry, but for PLDA the blessing is mixed -- length normalization Gaussianizes the between-speaker distribution, but violates the assumption of a speaker-independent within-speaker distribution. We propose PSDA, an analogue to PLDA that uses Von Mises-Fisher distributions on the hypersphere for both within and between-class distributions. We show how the self-conjugacy of this distribution gives closed-form likelihood-ratio scores, making it a drop-in replacement for PLDA at scoring time. All kinds of trials can be scored, including single-enroll and multi-enroll verification, as well as more complex likelihood-ratios that could be used in clustering and diarization. Learning is done via an EM-algorithm with closed-form updates. We explain the model and present some first experiments.

Dimitrios Koutsianos, Stavros Zacharopoulos, Yannis Panagakis et al.

This paper addresses source tracing in synthetic speech-identifying generative systems behind manipulated audio via speaker recognition-inspired pipelines. While prior work focuses on spoofing detection, source tracing lacks robust solutions. We evaluate two approaches: classification-based and metric-learning. We tested our methods on the MLAADv5 benchmark using ResNet and self-supervised learning (SSL) backbones. The results show that ResNet achieves competitive performance with the metric learning approach, matching and even exceeding SSL-based systems. Our work demonstrates ResNet's viability for source tracing while underscoring the need to optimize SSL representations for this task. Our work bridges speaker recognition methodologies with audio forensic challenges, offering new directions for combating synthetic media manipulation.

Dimitrios Koutsianos, Ladislav Mosner, Yannis Panagakis et al.

Performance in face and speaker verification is largely driven by margin based softmax losses like CosFace and ArcFace. Recently introduced $α$-divergence loss functions offer a compelling alternative, particularly for their ability to induce sparse solutions (when $α>1$). However, integrating an angular margin-crucial for verification tasks-is not straightforward. We find this integration can be achieved in at least two distinct ways: via the reference measure (prior probabilities) or via the logits (unnormalized log-likelihoods). In this paper, we explore both pathways, deriving two novel margin-based $α$-divergence losses: Q-Margin (margin in the reference measure) and A3M (margin in the logits). We identify and address a critical training instability in A3M-caused by the interplay of penalized logits and sparsity-with a simple yet effective prototype re-initialization strategy. Our methods achieve significant performance gains on the challenging IJB-B and IJB-C face verification benchmarks. We demonstrate similarly strong performance in speaker verification on VoxCeleb. Crucially, our models significantly outperform strong baselines at low false acceptance rates (FAR). This capability is crucial for practical high-security applications, such as banking authentication, when minimizing false authentications is paramount.

Vardis Georgilas, Themos Stafylakis

Medical dictation systems are essential tools in modern healthcare, enabling accurate and efficient conversion of speech into written medical documentation. The main objective of this paper is to create a domain-specific system for Greek medical speech transcriptions. The ultimate goal is to assist healthcare professionals by reducing the overload of manual documentation and improving workflow efficiency. Towards this goal, we develop a system that combines automatic speech recognition techniques with text correction model, allowing better handling of domain-specific terminology and linguistic variations in Greek. Our approach leverages both acoustic and textual modeling to create more realistic and reliable transcriptions. We focused on adapting existing language and speech technologies to the Greek medical context, addressing challenges such as complex medical terminology and linguistic inconsistencies. Through domain-specific fine-tuning, our system achieves more accurate and coherent transcriptions, contributing to the development of practical language technologies for the Greek healthcare sector.

Christos Vlachos, Nikolaos Stylianou, Alexandra Fiotaki et al.

We consider open-retrieval conversational question answering (OR-CONVQA), an extension of question answering where system responses need to be (i) aware of dialog history and (ii) grounded in documents (or document fragments) retrieved per question. Domain-specific OR-CONVQA training datasets are crucial for real-world applications, but hard to obtain. We propose a pipeline that capitalizes on the abundance of plain text documents in organizations (e.g., product documentation) to automatically produce realistic OR-CONVQA dialogs with annotations. Similarly to real-world humanannotated OR-CONVQA datasets, we generate in-dialog question-answer pairs, self-contained (decontextualized, e.g., no referring expressions) versions of user questions, and propositions (sentences expressing prominent information from the documents) the system responses are grounded in. We show how the synthetic dialogs can be used to train efficient question rewriters that decontextualize user questions, allowing existing dialog-unaware retrievers to be utilized. The retrieved information and the decontextualized question are then passed on to an LLM that generates the system's response.

Christos Vlachos, Themos Stafylakis, Ion Androutsopoulos

Creating effective and reliable task-oriented dialog systems (ToDSs) is challenging, not only because of the complex structure of these systems, but also due to the scarcity of training data, especially when several modules need to be trained separately, each one with its own input/output training examples. Data augmentation (DA), whereby synthetic training examples are added to the training data, has been successful in other NLP systems, but has not been explored as extensively in ToDSs. We empirically evaluate the effectiveness of DA methods in an end-to-end ToDS setting, where a single system is trained to handle all processing stages, from user inputs to system outputs. We experiment with two ToDSs (UBAR, GALAXY) on two datasets (MultiWOZ, KVRET). We consider three types of DA methods (word-level, sentence-level, dialog-level), comparing eight DA methods that have shown promising results in ToDSs and other NLP systems. We show that all DA methods considered are beneficial, and we highlight the best ones, also providing advice to practitioners. We also introduce a more challenging few-shot cross-domain ToDS setting, reaching similar conclusions.

Themos Stafylakis, Johan Rohdin, Lukas Burget

Speaker embeddings extracted with deep 2D convolutional neural networks are typically modeled as projections of first and second order statistics of channel-frequency pairs onto a linear layer, using either average or attentive pooling along the time axis. In this paper we examine an alternative pooling method, where pairwise correlations between channels for given frequencies are used as statistics. The method is inspired by style-transfer methods in computer vision, where the style of an image, modeled by the matrix of channel-wise correlations, is transferred to another image, in order to produce a new image having the style of the first and the content of the second. By drawing analogies between image style and speaker characteristics, and between image content and phonetic sequence, we explore the use of such channel-wise correlations features to train a ResNet architecture in an end-to-end fashion. Our experiments on VoxCeleb demonstrate the effectiveness of the proposed pooling method in speaker recognition.

Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis et al.

The goal of this work is to automatically determine whether and when a word of interest is spoken by a talking face, with or without the audio. We propose a zero-shot method suitable for in the wild videos. Our key contributions are: (1) a novel convolutional architecture, KWS-Net, that uses a similarity map intermediate representation to separate the task into (i) sequence matching, and (ii) pattern detection, to decide whether the word is there and when; (2) we demonstrate that if audio is available, visual keyword spotting improves the performance both for a clean and noisy audio signal. Finally, (3) we show that our method generalises to other languages, specifically French and German, and achieves a comparable performance to English with less language specific data, by fine-tuning the network pre-trained on English. The method exceeds the performance of the previous state-of-the-art visual keyword spotting architecture when trained and tested on the same benchmark, and also that of a state-of-the-art lip reading method.

Anna Silnova, Niko Brümmer, Johan Rohdin et al.

Speaker embeddings (x-vectors) extracted from very short segments of speech have recently been shown to give competitive performance in speaker diarization. We generalize this recipe by extracting from each speech segment, in parallel with the x-vector, also a diagonal precision matrix, thus providing a path for the propagation of information about the quality of the speech segment into a PLDA scoring backend. These precisions quantify the uncertainty about what the values of the embeddings might have been if they had been extracted from high quality speech segments. The proposed probabilistic embeddings (x-vectors with precisions) are interfaced with the PLDA model by treating the x-vectors as hidden variables and marginalizing them out. We apply the proposed probabilistic embeddings as input to an agglomerative hierarchical clustering (AHC) algorithm to do diarization in the DIHARD'19 evaluation set. We compute the full PLDA likelihood 'by the book' for each clustering hypothesis that is considered by AHC. We do joint discriminative training of the PLDA parameters and of the probabilistic x-vector extractor. We demonstrate accuracy gains relative to a baseline AHC algorithm, applied to traditional xvectors (without uncertainty), and which uses averaging of binary log-likelihood-ratios, rather than by-the-book scoring.

Hossein Zeinali, Themos Stafylakis, Georgia Athanasopoulou et al.

In this paper, we present the system description of the joint efforts of Brno University of Technology (BUT) and Omilia -- Conversational Intelligence for the ASVSpoof2019 Spoofing and Countermeasures Challenge. The primary submission for Physical access (PA) is a fusion of two VGG networks, trained on single and two-channels features. For Logical access (LA), our primary system is a fusion of VGG and the recently introduced SincNet architecture. The results on PA show that the proposed networks yield very competitive performance in all conditions and achieved 86\:\% relative improvement compared to the official baseline. On the other hand, the results on LA showed that although the proposed architecture and training strategy performs very well on certain spoofing attacks, it fails to generalize to certain attacks that are unseen during training.

Andreas Nautsch, Jose Patino, Amos Treiber et al.

In many voice biometrics applications there is a requirement to preserve privacy, not least because of the recently enforced General Data Protection Regulation (GDPR). Though progress in bringing privacy preservation to voice biometrics is lagging behind developments in other biometrics communities, recent years have seen rapid progress, with secure computation mechanisms such as homomorphic encryption being applied successfully to speaker recognition. Even so, the computational overhead incurred by processing speech data in the encrypted domain is substantial. While still tolerable for single biometric comparisons, most state-of-the-art systems perform some form of cohort-based score normalisation, requiring many thousands of biometric comparisons. The computational overhead is then prohibitive, meaning that one must accept either degraded performance (no score normalisation) or potential for privacy violations. This paper proposes the first computationally feasible approach to privacy-preserving cohort score normalisation. Our solution is a cohort pruning scheme based on secure multi-party computation which enables privacy-preserving score normalisation using probabilistic linear discriminant analysis (PLDA) comparisons. The solution operates upon binary voice representations. While the binarisation is lossy in biometric rank-1 performance, it supports computationally-feasible biometric rank-n comparisons in the encrypted domain.

Themos Stafylakis, Johan Rohdin, Oldrich Plchot et al.

Contrary to i-vectors, speaker embeddings such as x-vectors are incapable of leveraging unlabelled utterances, due to the classification loss over training speakers. In this paper, we explore an alternative training strategy to enable the use of unlabelled utterances in training. We propose to train speaker embedding extractors via reconstructing the frames of a target speech segment, given the inferred embedding of another speech segment of the same utterance. We do this by attaching to the standard speaker embedding extractor a decoder network, which we feed not merely with the speaker embedding, but also with the estimated phone sequence of the target frame sequence. The reconstruction loss can be used either as a single objective, or be combined with the standard speaker classification loss. In the latter case, it acts as a regularizer, encouraging generalizability to speakers unseen during training. In all cases, the proposed architectures are trained from scratch and in an end-to-end fashion. We demonstrate the benefits from the proposed approach on VoxCeleb and Speakers in the wild, and we report notable improvements over the baseline.

Johan Rohdin, Themos Stafylakis, Anna Silnova et al.

In this paper we investigate the use of adversarial domain adaptation for addressing the problem of language mismatch between speaker recognition corpora. In the context of speaker verification, adversarial domain adaptation methods aim at minimizing certain divergences between the distribution that the utterance-level features follow (i.e. speaker embeddings) when drawn from source and target domains (i.e. languages), while preserving their capacity in recognizing speakers. Neural architectures for extracting utterance-level representations enable us to apply adversarial adaptation methods in an end-to-end fashion and train the network jointly with the standard cross-entropy loss. We examine several configurations, such as the use of (pseudo-)labels on the target domain as well as domain labels in the feature extractor, and we demonstrate the effectiveness of our method on the challenging NIST SRE16 and SRE18 benchmarks.

Hossein Zeinali, Lukas Burget, Johan Rohdin et al.

Recently, speaker embeddings extracted with deep neural networks became the state-of-the-art method for speaker verification. In this paper we aim to facilitate its implementation on a more generic toolkit than Kaldi, which we anticipate to enable further improvements on the method. We examine several tricks in training, such as the effects of normalizing input features and pooled statistics, different methods for preventing overfitting as well as alternative non-linearities that can be used instead of Rectifier Linear Units. In addition, we investigate the difference in performance between TDNN and CNN, and between two types of attention mechanism. Experimental results on Speaker in the Wild, SRE 2016 and SRE 2018 datasets demonstrate the effectiveness of the proposed implementation.

Themos Stafylakis, Muhammad Haris Khan, Georgios Tzimiropoulos

Visual and audiovisual speech recognition are witnessing a renaissance which is largely due to the advent of deep learning methods. In this paper, we present a deep learning architecture for lipreading and audiovisual word recognition, which combines Residual Networks equipped with spatiotemporal input layers and Bidirectional LSTMs. The lipreading architecture attains 11.92% misclassification rate on the challenging Lipreading-In-The-Wild database, which is composed of excerpts from BBC-TV, each containing one of the 500 target words. Audiovisual experiments are performed using both intermediate and late integration, as well as several types and levels of environmental noise, and notable improvements over the audio-only network are reported, even in the case of clean speech. A further analysis on the utility of target word boundaries is provided, as well as on the capacity of the network in modeling the linguistic context of the target word. Finally, we examine difficult word pairs and discuss how visual information helps towards attaining higher recognition accuracy.

Stavros Petridis, Themos Stafylakis, Pingchuan Ma et al.

Recent works in speech recognition rely either on connectionist temporal classification (CTC) or sequence-to-sequence models for character-level recognition. CTC assumes conditional independence of individual characters, whereas attention-based models can provide nonsequential alignments. Therefore, we could use a CTC loss in combination with an attention-based model in order to force monotonic alignments and at the same time get rid of the conditional independence assumption. In this paper, we use the recently proposed hybrid CTC/attention architecture for audio-visual recognition of speech in-the-wild. To the best of our knowledge, this is the first time that such a hybrid architecture architecture is used for audio-visual recognition of speech. We use the LRS2 database and show that the proposed audio-visual model leads to an 1.3% absolute decrease in word error rate over the audio-only model and achieves the new state-of-the-art performance on LRS2 database (7% word error rate). We also observe that the audio-visual model significantly outperforms the audio-based model (up to 32.9% absolute improvement in word error rate) for several different types of noise as the signal-to-noise ratio decreases.

Themos Stafylakis, Georgios Tzimiropoulos

Visual keyword spotting (KWS) is the problem of estimating whether a text query occurs in a given recording using only video information. This paper focuses on visual KWS for words unseen during training, a real-world, practical setting which so far has received no attention by the community. To this end, we devise an end-to-end architecture comprising (a) a state-of-the-art visual feature extractor based on spatiotemporal Residual Networks, (b) a grapheme-to-phoneme model based on sequence-to-sequence neural networks, and (c) a stack of recurrent neural networks which learn how to correlate visual features with the keyword representation. Different to prior works on KWS, which try to learn word representations merely from sequences of graphemes (i.e. letters), we propose the use of a grapheme-to-phoneme encoder-decoder model which learns how to map words to their pronunciation. We demonstrate that our system obtains very promising visual-only KWS results on the challenging LRS2 database, for keywords unseen during training. We also show that our system outperforms a baseline which addresses KWS via automatic speech recognition (ASR), while it drastically improves over other recently proposed ASR-free KWS methods.

Niko Brummer, Anna Silnova, Lukas Burget et al.

Embeddings in machine learning are low-dimensional representations of complex input patterns, with the property that simple geometric operations like Euclidean distances and dot products can be used for classification and comparison tasks. The proposed meta-embeddings are special embeddings that live in more general inner product spaces. They are designed to propagate uncertainty to the final output in speaker recognition and similar applications. The familiar Gaussian PLDA model (GPLDA) can be re-formulated as an extractor for Gaussian meta-embeddings (GMEs), such that likelihood ratio scores are given by Hilbert space inner products between Gaussian likelihood functions. GMEs extracted by the GPLDA model have fixed precisions and do not propagate uncertainty. We show that a generalization to heavy-tailed PLDA gives GMEs with variable precisions, which do propagate uncertainty. Experiments on NIST SRE 2010 and 2016 show that the proposed method applied to i-vectors without length normalization is up to 20% more accurate than GPLDA applied to length-normalized ivectors.

Stavros Petridis, Themos Stafylakis, Pingchuan Ma et al.

Several end-to-end deep learning approaches have been recently presented which extract either audio or visual features from the input images or audio signals and perform speech recognition. However, research on end-to-end audiovisual models is very limited. In this work, we present an end-to-end audiovisual model based on residual networks and Bidirectional Gated Recurrent Units (BGRUs). To the best of our knowledge, this is the first audiovisual fusion model which simultaneously learns to extract features directly from the image pixels and audio waveforms and performs within-context word recognition on a large publicly available dataset (LRW). The model consists of two streams, one for each modality, which extract features directly from mouth regions and raw waveforms. The temporal dynamics in each stream/modality are modeled by a 2-layer BGRU and the fusion of multiple streams/modalities takes place via another 2-layer BGRU. A slight improvement in the classification rate over an end-to-end audio-only and MFCC-based model is reported in clean audio conditions and low levels of noise. In presence of high levels of noise, the end-to-end audiovisual model significantly outperforms both audio-only models.

Themos Stafylakis, Georgios Tzimiropoulos

In this paper we present a deep learning architecture for extracting word embeddings for visual speech recognition. The embeddings summarize the information of the mouth region that is relevant to the problem of word recognition, while suppressing other types of variability such as speaker, pose and illumination. The system is comprised of a spatiotemporal convolutional layer, a Residual Network and bidirectional LSTMs and is trained on the Lipreading in-the-wild database. We first show that the proposed architecture goes beyond state-of-the-art on closed-set word identification, by attaining 11.92% error rate on a vocabulary of 500 words. We then examine the capacity of the embeddings in modelling words unseen during training. We deploy Probabilistic Linear Discriminant Analysis (PLDA) to model the embeddings and perform low-shot learning experiments on words unseen during training. The experiments demonstrate that word-level visual speech recognition is feasible even in cases where the target words are not included in the training set.

Themos Stafylakis, Georgios Tzimiropoulos

We propose an end-to-end deep learning architecture for word-level visual speech recognition. The system is a combination of spatiotemporal convolutional, residual and bidirectional Long Short-Term Memory networks. We train and evaluate it on the Lipreading In-The-Wild benchmark, a challenging database of 500-size target-words consisting of 1.28sec video excerpts from BBC TV broadcasts. The proposed network attains word accuracy equal to 83.0, yielding 6.8 absolute improvement over the current state-of-the-art, without using information about word boundaries during training or testing.