Stavros Petridis

Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira et al.

We present RAVEn, a self-supervised multi-modal approach to jointly learn visual and auditory speech representations. Our pre-training objective involves encoding masked inputs, and then predicting contextualised targets generated by slowly-evolving momentum encoders. Driven by the inherent differences between video and audio, our design is asymmetric w.r.t. the two modalities' pretext tasks: Whereas the auditory stream predicts both the visual and auditory targets, the visual one predicts only the auditory targets. We observe strong results in low- and high-resource labelled data settings when fine-tuning the visual and auditory encoders resulting from a single pre-training stage, in which the encoders are jointly trained. Notably, RAVEn surpasses all self-supervised methods on visual speech recognition (VSR) on LRS3, and combining RAVEn with self-training using only 30 hours of labelled data even outperforms a recent semi-supervised method trained on 90,000 hours of non-public data. At the same time, we achieve state-of-the-art results in the LRS3 low-resource setting for auditory speech recognition (as well as for VSR). Our findings point to the viability of learning powerful speech representations entirely from raw video and audio, i.e., without relying on handcrafted features. Code and models are available at https://github.com/ahaliassos/raven.

Michał Stypułkowski, Konstantinos Vougioukas, Sen He et al.

Talking face generation has historically struggled to produce head movements and natural facial expressions without guidance from additional reference videos. Recent developments in diffusion-based generative models allow for more realistic and stable data synthesis and their performance on image and video generation has surpassed that of other generative models. In this work, we present an autoregressive diffusion model that requires only one identity image and audio sequence to generate a video of a realistic talking human head. Our solution is capable of hallucinating head movements, facial expressions, such as blinks, and preserving a given background. We evaluate our model on two different datasets, achieving state-of-the-art results on both of them.

Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez et al.

Audio-visual speech recognition has received a lot of attention due to its robustness against acoustic noise. Recently, the performance of automatic, visual, and audio-visual speech recognition (ASR, VSR, and AV-ASR, respectively) has been substantially improved, mainly due to the use of larger models and training sets. However, accurate labelling of datasets is time-consuming and expensive. Hence, in this work, we investigate the use of automatically-generated transcriptions of unlabelled datasets to increase the training set size. For this purpose, we use publicly-available pre-trained ASR models to automatically transcribe unlabelled datasets such as AVSpeech and VoxCeleb2. Then, we train ASR, VSR and AV-ASR models on the augmented training set, which consists of the LRS2 and LRS3 datasets as well as the additional automatically-transcribed data. We demonstrate that increasing the size of the training set, a recent trend in the literature, leads to reduced WER despite using noisy transcriptions. The proposed model achieves new state-of-the-art performance on AV-ASR on LRS2 and LRS3. In particular, it achieves a WER of 0.9% on LRS3, a relative improvement of 30% over the current state-of-the-art approach, and outperforms methods that have been trained on non-publicly available datasets with 26 times more training data.

Pingchuan Ma, Yujiang Wang, Stavros Petridis et al.

Several training strategies and temporal models have been recently proposed for isolated word lip-reading in a series of independent works. However, the potential of combining the best strategies and investigating the impact of each of them has not been explored. In this paper, we systematically investigate the performance of state-of-the-art data augmentation approaches, temporal models and other training strategies, like self-distillation and using word boundary indicators. Our results show that Time Masking (TM) is the most important augmentation followed by mixup and Densely-Connected Temporal Convolutional Networks (DC-TCN) are the best temporal model for lip-reading of isolated words. Using self-distillation and word boundary indicators is also beneficial but to a lesser extent. A combination of all the above methods results in a classification accuracy of 93.4%, which is an absolute improvement of 4.6% over the current state-of-the-art performance on the LRW dataset. The performance can be further improved to 94.1% by pre-training on additional datasets. An error analysis of the various training strategies reveals that the performance improves by increasing the classification accuracy of hard-to-recognise words.

Rodrigo Mira, Alexandros Haliassos, Stavros Petridis et al.

Video-to-speech synthesis (also known as lip-to-speech) refers to the translation of silent lip movements into the corresponding audio. This task has received an increasing amount of attention due to its self-supervised nature (i.e., can be trained without manual labelling) combined with the ever-growing collection of audio-visual data available online. Despite these strong motivations, contemporary video-to-speech works focus mainly on small- to medium-sized corpora with substantial constraints in both vocabulary and setting. In this work, we introduce a scalable video-to-speech framework consisting of two components: a video-to-spectrogram predictor and a pre-trained neural vocoder, which converts the mel-frequency spectrograms into waveform audio. We achieve state-of-the art results for GRID and considerably outperform previous approaches on LRW. More importantly, by focusing on spectrogram prediction using a simple feedforward model, we can efficiently and effectively scale our method to very large and unconstrained datasets: To the best of our knowledge, we are the first to show intelligible results on the challenging LRS3 dataset.

Xubo Liu, Egor Lakomkin, Konstantinos Vougioukas et al.

Recently reported state-of-the-art results in visual speech recognition (VSR) often rely on increasingly large amounts of video data, while the publicly available transcribed video datasets are limited in size. In this paper, for the first time, we study the potential of leveraging synthetic visual data for VSR. Our method, termed SynthVSR, substantially improves the performance of VSR systems with synthetic lip movements. The key idea behind SynthVSR is to leverage a speech-driven lip animation model that generates lip movements conditioned on the input speech. The speech-driven lip animation model is trained on an unlabeled audio-visual dataset and could be further optimized towards a pre-trained VSR model when labeled videos are available. As plenty of transcribed acoustic data and face images are available, we are able to generate large-scale synthetic data using the proposed lip animation model for semi-supervised VSR training. We evaluate the performance of our approach on the largest public VSR benchmark - Lip Reading Sentences 3 (LRS3). SynthVSR achieves a WER of 43.3% with only 30 hours of real labeled data, outperforming off-the-shelf approaches using thousands of hours of video. The WER is further reduced to 27.9% when using all 438 hours of labeled data from LRS3, which is on par with the state-of-the-art self-supervised AV-HuBERT method. Furthermore, when combined with large-scale pseudo-labeled audio-visual data SynthVSR yields a new state-of-the-art VSR WER of 16.9% using publicly available data only, surpassing the recent state-of-the-art approaches trained with 29 times more non-public machine-transcribed video data (90,000 hours). Finally, we perform extensive ablation studies to understand the effect of each component in our proposed method.

Rodrigo Mira, Buye Xu, Jacob Donley et al.

Audio-visual speech enhancement aims to extract clean speech from a noisy environment by leveraging not only the audio itself but also the target speaker's lip movements. This approach has been shown to yield improvements over audio-only speech enhancement, particularly for the removal of interfering speech. Despite recent advances in speech synthesis, most audio-visual approaches continue to use spectral mapping/masking to reproduce the clean audio, often resulting in visual backbones added to existing speech enhancement architectures. In this work, we propose LA-VocE, a new two-stage approach that predicts mel-spectrograms from noisy audio-visual speech via a transformer-based architecture, and then converts them into waveform audio using a neural vocoder (HiFi-GAN). We train and evaluate our framework on thousands of speakers and 11+ different languages, and study our model's ability to adapt to different levels of background noise and speech interference. Our experiments show that LA-VocE outperforms existing methods according to multiple metrics, particularly under very noisy scenarios.

Andreas Zinonos, Alexandros Haliassos, Pingchuan Ma et al.

Cross-lingual self-supervised learning has been a growing research topic in the last few years. However, current works only explored the use of audio signals to create representations. In this work, we study cross-lingual self-supervised visual representation learning. We use the recently-proposed Raw Audio-Visual Speech Encoders (RAVEn) framework to pre-train an audio-visual model with unlabelled multilingual data, and then fine-tune the visual model on labelled transcriptions. Our experiments show that: (1) multi-lingual models with more data outperform monolingual ones, but, when keeping the amount of data fixed, monolingual models tend to reach better performance; (2) multi-lingual outperforms English-only pre-training; (3) using languages which are more similar yields better results; and (4) fine-tuning on unseen languages is competitive to using the target language in the pre-training set. We hope our study inspires future research on non-English-only speech representation learning.

Adriana Fernandez-Lopez, Honglie Chen, Pingchuan Ma et al.

Recent advances in deep neural networks have achieved unprecedented success in visual speech recognition. However, there remains substantial disparity between current methods and their deployment in resource-constrained devices. In this work, we explore different magnitude-based pruning techniques to generate a lightweight model that achieves higher performance than its dense model equivalent, especially under the presence of visual noise. Our sparse models achieve state-of-the-art results at 10% sparsity on the LRS3 dataset and outperform the dense equivalent up to 70% sparsity. We evaluate our 50% sparse model on 7 different visual noise types and achieve an overall absolute improvement of more than 2% WER compared to the dense equivalent. Our results confirm that sparse networks are more resistant to noise than dense networks.

Marija Jegorova, Stavros Petridis, Maja Pantic

This work focuses on the apparent emotional reaction recognition (AERR) from the video-only input, conducted in a self-supervised fashion. The network is first pre-trained on different self-supervised pretext tasks and later fine-tuned on the downstream target task. Self-supervised learning facilitates the use of pre-trained architectures and larger datasets that might be deemed unfit for the target task and yet might be useful to learn informative representations and hence provide useful initializations for further fine-tuning on smaller more suitable data. Our presented contribution is two-fold: (1) an analysis of different state-of-the-art (SOTA) pretext tasks for the video-only apparent emotional reaction recognition architecture, and (2) an analysis of various combinations of the regression and classification losses that are likely to improve the performance further. Together these two contributions result in the current state-of-the-art performance for the video-only spontaneous apparent emotional reaction recognition with continuous annotations.

Yujiang Wang, Mingzhi Dong, Jie Shen et al.

This paper presents a novel method for face clustering in videos using a video-centralised transformer. Previous works often employed contrastive learning to learn frame-level representation and used average pooling to aggregate the features along the temporal dimension. This approach may not fully capture the complicated video dynamics. In addition, despite the recent progress in video-based contrastive learning, few have attempted to learn a self-supervised clustering-friendly face representation that benefits the video face clustering task. To overcome these limitations, our method employs a transformer to directly learn video-level representations that can better reflect the temporally-varying property of faces in videos, while we also propose a video-centralised self-supervised framework to train the transformer model. We also investigate face clustering in egocentric videos, a fast-emerging field that has not been studied yet in works related to face clustering. To this end, we present and release the first large-scale egocentric video face clustering dataset named EasyCom-Clustering. We evaluate our proposed method on both the widely used Big Bang Theory (BBT) dataset and the new EasyCom-Clustering dataset. Results show the performance of our video-centralised transformer has surpassed all previous state-of-the-art methods on both benchmarks, exhibiting a self-attentive understanding of face videos.

Umberto Cappellazzo, Minsu Kim, Honglie Chen et al.

Multimodal large language models (MLLMs) have recently become a focal point of research due to their formidable multimodal understanding capabilities. For example, in the audio and speech domains, an LLM can be equipped with (automatic) speech recognition (ASR) abilities by just concatenating the audio tokens, computed with an audio encoder, and the text tokens to achieve state-of-the-art results. On the contrary, tasks like visual and audio-visual speech recognition (VSR/AVSR), which also exploit noise-invariant lip movement information, have received little or no attention. To bridge this gap, we propose Llama-AVSR, a new MLLM with strong audio-visual speech recognition capabilities. It leverages pre-trained audio and video encoders to produce modality-specific tokens which, together with the text tokens, are processed by a pre-trained LLM (e.g., Llama3.1-8B) to yield the resulting response in an auto-regressive fashion. Llama-AVSR requires a small number of trainable parameters as only modality-specific projectors and LoRA modules are trained whereas the multi-modal encoders and LLM are kept frozen. We evaluate our proposed approach on LRS3, the largest public AVSR benchmark, and we achieve new state-of-the-art results for the tasks of ASR and AVSR with a WER of 0.79% and 0.77%, respectively. To bolster our results, we investigate the key factors that underpin the effectiveness of Llama-AVSR: the choice of the pre-trained encoders and LLM, the efficient integration of LoRA modules, and the optimal performance-efficiency trade-off obtained via modality-aware compression rates.

Pingchuan Ma, Niko Moritz, Stavros Petridis et al.

In this work, we propose a streaming AV-ASR system based on a hybrid connectionist temporal classification (CTC)/attention neural network architecture. The audio and the visual encoder neural networks are both based on the conformer architecture, which is made streamable using chunk-wise self-attention (CSA) and causal convolution. Streaming recognition with a decoder neural network is realized by using the triggered attention technique, which performs time-synchronous decoding with joint CTC/attention scoring. Additionally, we propose a novel alignment regularization technique that promotes synchronization of the audio and visual encoder, which in turn results in better word error rates (WERs) at all SNR levels for streaming and offline AV-ASR models. The proposed AV-ASR model achieves WERs of 2.0% and 2.6% on the Lip Reading Sentences 3 (LRS3) dataset in an offline and online setup, respectively, which both present state-of-the-art results when no external training data are used.

Honglie Chen, Rodrigo Mira, Stavros Petridis et al.

In this paper, we aim to generate clean speech frame by frame from a live video stream and a noisy audio stream without relying on future inputs. To this end, we propose RT-LA-VocE, which completely re-designs every component of LA-VocE, a state-of-the-art non-causal audio-visual speech enhancement model, to perform causal real-time inference with a 40ms input frame. We do so by devising new visual and audio encoders that rely solely on past frames, replacing the Transformer encoder with the Emformer, and designing a new causal neural vocoder C-HiFi-GAN. On the popular AVSpeech dataset, we show that our algorithm achieves state-of-the-art results in all real-time scenarios. More importantly, each component is carefully tuned to minimize the algorithm latency to the theoretical minimum (40ms) while maintaining a low end-to-end processing latency of 28.15ms per frame, enabling real-time frame-by-frame enhancement with minimal delay.

Umberto Cappellazzo, Stavros Petridis, Maja Pantic

Audio-Visual Speech Recognition (AVSR) leverages both acoustic and visual information for robust recognition under noise. However, how models balance these modalities remains unclear. We present Dr. SHAP-AV, a framework using Shapley values to analyze modality contributions in AVSR. Through experiments on six models across two benchmarks and varying SNR levels, we introduce three analyses: Global SHAP for overall modality balance, Generative SHAP for contribution dynamics during decoding, and Temporal Alignment SHAP for input-output correspondence. Our findings reveal that models shift toward visual reliance under noise yet maintain high audio contributions even under severe degradation. Modality balance evolves during generation, temporal alignment holds under noise, and SNR is the dominant factor driving modality weighting. These findings expose a persistent audio bias, motivating ad-hoc modality-weighting mechanisms and Shapley-based attribution as a standard AVSR diagnostic.

Andreas Zinonos, Michał Stypułkowski, Antoni Bigata et al.

We present FlashLips, a two-stage, mask-free lip-sync system that decouples lips control from rendering and achieves real-time performance running at over 100 FPS on a single GPU, while matching the visual quality of larger state-of-the-art models. Stage 1 is a compact, one-step latent-space editor that reconstructs an image using a reference identity, a masked target frame, and a low-dimensional lips-pose vector, trained purely with reconstruction losses - no GANs or diffusion. To remove explicit masks at inference, we use self-supervision: we generate mouth-altered variants of the target image, that serve as pseudo ground truth for fine-tuning, teaching the network to localize edits to the lips while preserving the rest. Stage 2 is an audio-to-pose transformer trained with a flow-matching objective to predict lips-poses vectors from speech. Together, these stages form a simple and stable pipeline that combines deterministic reconstruction with robust audio control, delivering high perceptual quality and faster-than-real-time speed.

Alexandros Haliassos, Rodrigo Mira, Stavros Petridis

Unified Speech Recognition (USR) has emerged as a semi-supervised framework for training a single model for audio, visual, and audiovisual speech recognition, achieving state-of-the-art results on in-distribution benchmarks. However, its reliance on autoregressive pseudo-labelling makes training expensive, while its decoupled supervision of CTC and attention branches increases susceptibility to self-reinforcing errors, particularly under distribution shifts involving longer sequences, noise, or unseen domains. We propose CTC-driven teacher forcing, where greedily decoded CTC pseudo-labels are fed into the decoder to generate attention targets in a single forward pass. Although these can be globally incoherent, in the pseudo-labelling setting they enable efficient and effective knowledge transfer. Because CTC and CTC-driven attention pseudo-labels have the same length, the decoder can predict both simultaneously, benefiting from the robustness of CTC and the expressiveness of attention without costly beam search. We further propose mixed sampling to mitigate the exposure bias of the decoder relying solely on CTC inputs. The resulting method, USR 2.0, halves training time, improves robustness to out-of-distribution inputs, and achieves state-of-the-art results on LRS3, LRS2, and WildVSR, surpassing USR and modality-specific self-supervised baselines.

Umberto Cappellazzo, Xubo Liu, Pingchuan Ma et al.

Large language models (LLMs) have recently achieved impressive results in speech recognition across multiple modalities, including Auditory Speech Recognition (ASR), Visual Speech Recognition (VSR), and Audio-Visual Speech Recognition (AVSR). Despite this progress, current LLM-based approaches typically address each task independently, training separate models that raise computational and deployment resource use while missing potential cross-task synergies. They also rely on fixed-rate token compression, which restricts flexibility in balancing accuracy with efficiency. These limitations highlight the need for a unified framework that can support ASR, VSR, and AVSR while enabling elastic inference. To this end, we present Omni-AVSR, a unified audio-visual LLM that combines efficient multi-granularity training with parameter-efficient adaptation. Specifically, we adapt the matryoshka representation learning paradigm to efficiently train across multiple audio and visual granularities, reducing its inherent training resource use. Furthermore, we explore three LoRA-based strategies for adapting the backbone LLM, balancing shared and task-specific specialization. Experiments on LRS2 and LRS3 show that Omni-AVSR achieves comparable or superior accuracy to state-of-the-art baselines while training a single model at substantially lower training and deployment resource use. The model also remains robust under acoustic noise, and we analyze its scaling behavior as LLM size increases, providing insights into the trade-off between performance and efficiency.

Alexandros Haliassos, Rodrigo Mira, Honglie Chen et al.

Research in auditory, visual, and audiovisual speech recognition (ASR, VSR, and AVSR, respectively) has traditionally been conducted independently. Even recent self-supervised studies addressing two or all three tasks simultaneously tend to yield separate models, leading to disjoint inference pipelines with increased memory requirements and redundancies. This paper proposes unified training strategies for these systems. We demonstrate that training a single model for all three tasks enhances VSR and AVSR performance, overcoming typical optimisation challenges when training from scratch. Moreover, we introduce a greedy pseudo-labelling approach to more effectively leverage unlabelled samples, addressing shortcomings in related self-supervised methods. Finally, we develop a self-supervised pre-training method within our framework, proving its effectiveness alongside our semi-supervised approach. Despite using a single model for all tasks, our unified approach achieves state-of-the-art performance compared to recent methods on LRS3 and LRS2 for ASR, VSR, and AVSR, as well as on the newly released WildVSR dataset. Code and models are available at https://github.com/ahaliassos/usr.

Minsu Kim, Rodrigo Mira, Honglie Chen et al.

In this paper, we investigate a novel approach for Target Speech Extraction (TSE), which relies solely on textual context to extract the target speech. We refer to this task as Contextual Speech Extraction (CSE). Unlike traditional TSE methods that rely on pre-recorded enrollment utterances, video of the target speaker's face, spatial information, or other explicit cues to identify the target stream, our proposed method requires only a few turns of previous dialogue (or monologue) history. This approach is naturally feasible in mobile messaging environments where voice recordings are typically preceded by textual dialogue that can be leveraged implicitly. We present three CSE models and analyze their performances on three datasets. Through our experiments, we demonstrate that even when the model relies purely on dialogue history, it can achieve over 90 % accuracy in identifying the correct target stream with only two previous dialogue turns. Furthermore, we show that by leveraging both textual context and enrollment utterances as cues during training, we further enhance our model's flexibility and effectiveness, allowing us to use either cue during inference, or combine both for improved performance. Samples and code available on https://miraodasilva.github.io/cse-project-page .

Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis et al.

Although current deep learning-based face forgery detectors achieve impressive performance in constrained scenarios, they are vulnerable to samples created by unseen manipulation methods. Some recent works show improvements in generalisation but rely on cues that are easily corrupted by common post-processing operations such as compression. In this paper, we propose LipForensics, a detection approach capable of both generalising to novel manipulations and withstanding various distortions. LipForensics targets high-level semantic irregularities in mouth movements, which are common in many generated videos. It consists in first pretraining a spatio-temporal network to perform visual speech recognition (lipreading), thus learning rich internal representations related to natural mouth motion. A temporal network is subsequently finetuned on fixed mouth embeddings of real and forged data in order to detect fake videos based on mouth movements without overfitting to low-level, manipulation-specific artefacts. Extensive experiments show that this simple approach significantly surpasses the state-of-the-art in terms of generalisation to unseen manipulations and robustness to perturbations, as well as shed light on the factors responsible for its performance. Code is available on GitHub.

Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas et al.

Head avatars animated by visual signals have gained popularity, particularly in cross-driving synthesis where the driver differs from the animated character, a challenging but highly practical approach. The recently presented MegaPortraits model has demonstrated state-of-the-art results in this domain. We conduct a deep examination and evaluation of this model, with a particular focus on its latent space for facial expression descriptors, and uncover several limitations with its ability to express intense face motions. To address these limitations, we propose substantial changes in both training pipeline and model architecture, to introduce our EMOPortraits model, where we: Enhance the model's capability to faithfully support intense, asymmetric face expressions, setting a new state-of-the-art result in the emotion transfer task, surpassing previous methods in both metrics and quality. Incorporate speech-driven mode to our model, achieving top-tier performance in audio-driven facial animation, making it possible to drive source identity through diverse modalities, including visual signal, audio, or a blend of both. We propose a novel multi-view video dataset featuring a wide range of intense and asymmetric facial expressions, filling the gap with absence of such data in existing datasets.

Alexandros Haliassos, Andreas Zinonos, Rodrigo Mira et al.

Self-supervision has recently shown great promise for learning visual and auditory speech representations from unlabelled data. In this work, we propose BRAVEn, an extension to the recent RAVEn method, which learns speech representations entirely from raw audio-visual data. Our modifications to RAVEn enable BRAVEn to achieve state-of-the-art results among self-supervised methods in various settings. Moreover, we observe favourable scaling behaviour by increasing the amount of unlabelled data well beyond other self-supervised works. In particular, we achieve 20.0% / 1.7% word error rate for VSR / ASR on the LRS3 test set, with only 30 hours of labelled data and no external ASR models. Our results suggest that readily available unlabelled audio-visual data can largely replace costly transcribed data.

Umberto Cappellazzo, Minsu Kim, Stavros Petridis

Audio-Visual Speech Recognition (AVSR) leverages audio and visual modalities to improve robustness in noisy environments. Recent advances in Large Language Models (LLMs) show strong performance in speech recognition, including AVSR. However, the long speech representations lead to high computational costs for LLMs. Prior methods compress inputs before feeding them to LLMs, but high compression often harms accuracy. To address this, we propose Llama-MTSK, the first Matryoshka-based Multimodal LLM for AVSR, which flexibly adapts audio-visual token allocation under varying compute constraints. Inspired by Matryoshka Representation Learning, our model encodes representations at multiple granularities with a single architecture, avoiding the need for separate models. For efficient fine-tuning, we introduce three LoRA-based strategies using global and scale-specific modules. Evaluations on major AVSR datasets show Llama-MTSK matches or outperforms models trained at fixed compression levels.

Umberto Cappellazzo, Minsu Kim, Stavros Petridis et al.

Audio-Visual Speech Recognition (AVSR) enhances robustness in noisy environments by integrating visual cues. While recent advances integrate Large Language Models (LLMs) into AVSR, their high computational cost hinders deployment in resource-constrained settings. To address this, we propose Llama-SMoP, an efficient Multimodal LLM that employs a Sparse Mixture of Projectors (SMoP) module to scale model capacity without increasing inference costs. By incorporating sparsely-gated mixture-of-experts (MoE) projectors, Llama-SMoP enables the use of smaller LLMs while maintaining strong performance. We explore three SMoP configurations and show that Llama-SMoP DEDR (Disjoint-Experts, Disjoint-Routers), which uses modality-specific routers and experts, achieves superior performance on ASR, VSR, and AVSR tasks. Ablation studies confirm its effectiveness in expert activation, scalability, and noise robustness.

Jeong Hun Yeo, Minsu Kim, Chae Won Kim et al.

We explore a novel zero-shot Audio-Visual Speech Recognition (AVSR) framework, dubbed Zero-AVSR, which enables speech recognition in target languages without requiring any audio-visual speech data in those languages. Specifically, we introduce the Audio-Visual Speech Romanizer (AV-Romanizer), which learns language-agnostic speech representations by predicting Roman text. Then, by leveraging the strong multilingual modeling capabilities of Large Language Models (LLMs), we propose converting the predicted Roman text into language-specific graphemes, forming the proposed Cascaded Zero-AVSR. Taking it a step further, we explore a unified Zero-AVSR approach by directly integrating the audio-visual speech representations encoded by the AV-Romanizer into the LLM. This is achieved through finetuning the adapter and the LLM using our proposed multi-task learning scheme. To capture the wide spectrum of phonetic and linguistic diversity, we also introduce a Multilingual Audio-Visual Romanized Corpus (MARC) consisting of 2,916 hours of audio-visual speech data across 82 languages, along with transcriptions in both language-specific graphemes and Roman text. Extensive analysis and experiments confirm that the proposed Zero-AVSR framework has the potential to expand language support beyond the languages seen during the training of the AV-Romanizer.

Yufeng Yin, Ishwarya Ananthabhotla, Vamsi Krishna Ithapu et al.

Individuals with impaired hearing experience difficulty in conversations, especially in noisy environments. This difficulty often manifests as a change in behavior and may be captured via facial expressions, such as the expression of discomfort or fatigue. In this work, we build on this idea and introduce the problem of detecting hearing loss from an individual's facial expressions during a conversation. Building machine learning models that can represent hearing-related facial expression changes is a challenge. In addition, models need to disentangle spurious age-related correlations from hearing-driven expressions. To this end, we propose a self-supervised pre-training strategy tailored for the modeling of expression variations. We also use adversarial representation learning to mitigate the age bias. We evaluate our approach on a large-scale egocentric dataset with real-world conversational scenarios involving subjects with hearing loss and show that our method for hearing loss detection achieves superior performance over baselines.

Antoni Bigata, Michał Stypułkowski, Rodrigo Mira et al.

Current audio-driven facial animation methods achieve impressive results for short videos but suffer from error accumulation and identity drift when extended to longer durations. Existing methods attempt to mitigate this through external spatial control, increasing long-term consistency but compromising the naturalness of motion. We propose KeyFace, a novel two-stage diffusion-based framework, to address these issues. In the first stage, keyframes are generated at a low frame rate, conditioned on audio input and an identity frame, to capture essential facial expressions and movements over extended periods of time. In the second stage, an interpolation model fills in the gaps between keyframes, ensuring smooth transitions and temporal coherence. To further enhance realism, we incorporate continuous emotion representations and handle a wide range of non-speech vocalizations (NSVs), such as laughter and sighs. We also introduce two new evaluation metrics for assessing lip synchronization and NSV generation. Experimental results show that KeyFace outperforms state-of-the-art methods in generating natural, coherent facial animations over extended durations, successfully encompassing NSVs and continuous emotions.

Antoni Bigata, Rodrigo Mira, Stella Bounareli et al.

Lip synchronization, known as the task of aligning lip movements in an existing video with new input audio, is typically framed as a simpler variant of audio-driven facial animation. However, as well as suffering from the usual issues in talking head generation (e.g., temporal consistency), lip synchronization presents significant new challenges such as expression leakage from the input video and facial occlusions, which can severely impact real-world applications like automated dubbing, but are often neglected in existing works. To address these shortcomings, we present KeySync, a two-stage framework that succeeds in solving the issue of temporal consistency, while also incorporating solutions for leakage and occlusions using a carefully designed masking strategy. We show that KeySync achieves state-of-the-art results in lip reconstruction and cross-synchronization, improving visual quality and reducing expression leakage according to LipLeak, our novel leakage metric. Furthermore, we demonstrate the effectiveness of our new masking approach in handling occlusions and validate our architectural choices through several ablation studies. Code and model weights can be found at https://antonibigata.github.io/KeySync.

Minsu Kim, Pingchuan Ma, Honglie Chen et al.

This paper explores multi-modal controllable Text-to-Speech Synthesis (TTS) where the voice can be generated from face image, and the characteristics of output speech (e.g., pace, noise level, distance, tone, place) can be controllable with natural text description. Specifically, we aim to mitigate the following three challenges in face-driven TTS systems. 1) To overcome the limited audio quality of audio-visual speech corpora, we propose a training method that additionally utilizes high-quality audio-only speech corpora. 2) To generate voices not only from real human faces but also from artistic portraits, we propose augmenting the input face image with stylization. 3) To consider one-to-many possibilities in face-to-voice mapping and ensure consistent voice generation at the same time, we propose to first employ sampling-based decoding and then use prompting with generated speech samples. Experimental results validate the proposed model's effectiveness in face-driven voice synthesis.

Kazuaki Mishima, Antoni Bigata Casademunt, Stavros Petridis et al.

Human facial images encode a rich spectrum of information, encompassing both stable identity-related traits and mutable attributes such as pose, expression, and emotion. While recent advances in image generation have enabled high-quality identity-conditional face synthesis, precise control over non-identity attributes remains challenging, and disentangling identity from these mutable factors is particularly difficult. To address these limitations, we propose a novel identity-conditional diffusion model that introduces two lightweight control modules designed to independently manipulate facial pose, expression, and emotion without compromising identity preservation. These modules are embedded within the cross-attention layers of the base diffusion model, enabling precise attribute control with minimal parameter overhead. Furthermore, our tailored training strategy, which leverages cross-attention between the identity feature and each non-identity control feature, encourages identity features to remain orthogonal to control signals, enhancing controllability and diversity. Quantitative and qualitative evaluations, along with perceptual user studies, demonstrate that our method surpasses existing approaches in terms of control accuracy over pose, expression, and emotion, while also improving generative diversity under identity-only conditioning.

Anand, Umberto Cappellazzo, Stavros Petridis et al.

Large language models (LLMs) have recently advanced auditory speech recognition (ASR), visual speech recognition (VSR), and audio-visual speech recognition (AVSR). However, understanding of their internal dynamics under fine-tuning remains limited. In natural language processing, recent work has revealed attention sinks, tokens that attract disproportionately high attention, and associated massive activations in which some features of sink tokens exhibit huge activation in LLMs. In this work, we are the first to study these phenomena in multimodal speech recognition. Through a detailed analysis of audio-visual LLMs, we identify attention sinks and massive activations not only at the BOS token but also at intermediate low-semantic tokens across ASR, VSR, and AVSR. We show that massive activations originate in the MLP layers and correspond to fixed feature indices across all sink tokens. We further show that intermediate sink tokens exhibit high cosine similarity to the BOS token, thereby amplifying attention and activation. Building on these insights, we introduce a simple decorrelation loss that reduces cosine similarity between BOS and other tokens, effectively mitigating intermediate sinks and massive activations. Furthermore, our method improves word error rate (WER) under high audio-visual feature downsampling while remaining stable at lower downsampling rates.

Umberto Cappellazzo, Minsu Kim, Pingchuan Ma et al.

Large language models (LLMs) have recently shown strong potential in audio-visual speech recognition (AVSR), but their high computational demands and sensitivity to token granularity limit their practicality in resource-constrained settings. Token compression methods can reduce inference cost, but they require fixing a compression rate in advance and produce a single fixed-length output, offering no flexibility to balance information density and efficiency at inference time. Matryoshka representation learning (MRL) addresses this by enabling a single model to operate across multiple token granularities, allowing compression rates to be adjusted dynamically. However, current MRL-based methods treat each scale independently during training, limiting cross-scale generalization, robustness at high compression, and interpretability. To overcome these limitations, we propose MoME (Mixture of Matryoshka Experts), a novel framework that integrates sparse Mixture-of-Experts (MoE) into MRL-based LLMs for AVSR. MoME augments a frozen LLM with top-k routed and shared experts, allowing dynamic capacity allocation across scales and modalities. A shared router promotes consistent expert activation across granularities, enabling compressed sequences to benefit from representations learned at lower compression. Experiments on LRS2 and LRS3 demonstrate that MoME achieves state-of-the-art performance across AVSR, ASR, and VSR tasks, while requiring significantly fewer parameters and maintaining robustness under noise. MoME unifies the adaptability of MRL with the efficiency of MoE, offering a scalable and interpretable solution for resource-aware speech recognition.

Qiao Xiao, Pingchuan Ma, Adriana Fernandez-Lopez et al.

The recent success of Automatic Speech Recognition (ASR) is largely attributed to the ever-growing amount of training data. However, this trend has made model training prohibitively costly and imposed computational demands. While data pruning has been proposed to mitigate this issue by identifying a small subset of relevant data, its application in ASR has been barely explored, and existing works often entail significant overhead to achieve meaningful results. To fill this gap, this paper presents the first investigation of dynamic data pruning for ASR, finding that we can reach the full-data performance by dynamically selecting 70% of data. Furthermore, we introduce Dynamic Data Pruning for ASR (DDP-ASR), which offers several fine-grained pruning granularities specifically tailored for speech-related datasets, going beyond the conventional pruning of entire time sequences. Our intensive experiments show that DDP-ASR can save up to 1.6x training time with negligible performance loss.

Adriana Fernandez-Lopez, Honglie Chen, Pingchuan Ma et al.

Pre-trained models have been a foundational approach in speech recognition, albeit with associated additional costs. In this study, we propose a regularization technique that facilitates the training of visual and audio-visual speech recognition models (VSR and AVSR) from scratch. This approach, abbreviated as \textbf{MSRS} (Multimodal Speech Recognition from Scratch), introduces a sparse regularization that rapidly learns sparse structures within the dense model at the very beginning of training, which receives healthier gradient flow than the dense equivalent. Once the sparse mask stabilizes, our method allows transitioning to a dense model or keeping a sparse model by updating non-zero values. MSRS achieves competitive results in VSR and AVSR with 21.1% and 0.9% WER on the LRS3 benchmark, while reducing training time by at least 2x. We explore other sparse approaches and show that only MSRS enables training from scratch by implicitly masking the weights affected by vanishing gradients.

Antoni Bigata Casademunt, Rodrigo Mira, Nikita Drobyshev et al.

Speech-driven animation has gained significant traction in recent years, with current methods achieving near-photorealistic results. However, the field remains underexplored regarding non-verbal communication despite evidence demonstrating its importance in human interaction. In particular, generating laughter sequences presents a unique challenge due to the intricacy and nuances of this behaviour. This paper aims to bridge this gap by proposing a novel model capable of generating realistic laughter sequences, given a still portrait and an audio clip containing laughter. We highlight the failure cases of traditional facial animation methods and leverage recent advances in diffusion models to produce convincing laughter videos. We train our model on a diverse set of laughter datasets and introduce an evaluation metric specifically designed for laughter. When compared with previous speech-driven approaches, our model achieves state-of-the-art performance across all metrics, even when these are re-trained for laughter generation. Our code and project are publicly available

Yujiang Wang, Anshul Thakur, Mingzhi Dong et al.

The prevalence of artificial intelligence (AI) has envisioned an era of healthcare democratisation that promises every stakeholder a new and better way of life. However, the advancement of clinical AI research is significantly hurdled by the dearth of data democratisation in healthcare. To truly democratise data for AI studies, challenges are two-fold: 1. the sensitive information in clinical data should be anonymised appropriately, and 2. AI-oriented clinical knowledge should flow freely across organisations. This paper considers a recent deep-learning advent, dataset condensation (DC), as a stone that kills two birds in democratising healthcare data. The condensed data after DC, which can be viewed as statistical metadata, abstracts original clinical records and irreversibly conceals sensitive information at individual levels; nevertheless, it still preserves adequate knowledge for learning deep neural networks (DNNs). More favourably, the compressed volumes and the accelerated model learnings of condensed data portray a more efficient clinical knowledge sharing and flowing system, as necessitated by data democratisation. We underline DC's prospects for democratising clinical data, specifically electrical healthcare records (EHRs), for AI research through experimental results and analysis across three healthcare datasets of varying data types.

Pingchuan Ma, Stavros Petridis, Maja Pantic

Visual speech recognition (VSR) aims to recognize the content of speech based on lip movements, without relying on the audio stream. Advances in deep learning and the availability of large audio-visual datasets have led to the development of much more accurate and robust VSR models than ever before. However, these advances are usually due to the larger training sets rather than the model design. Here we demonstrate that designing better models is equally as important as using larger training sets. We propose the addition of prediction-based auxiliary tasks to a VSR model, and highlight the importance of hyperparameter optimization and appropriate data augmentations. We show that such a model works for different languages and outperforms all previous methods trained on publicly available datasets by a large margin. It even outperforms models that were trained on non-publicly available datasets containing up to to 21 times more data. We show, furthermore, that using additional training data, even in other languages or with automatically generated transcriptions, results in further improvement.

Alexandros Haliassos, Rodrigo Mira, Stavros Petridis et al.

One of the most pressing challenges for the detection of face-manipulated videos is generalising to forgery methods not seen during training while remaining effective under common corruptions such as compression. In this paper, we examine whether we can tackle this issue by harnessing videos of real talking faces, which contain rich information on natural facial appearance and behaviour and are readily available in large quantities online. Our method, termed RealForensics, consists of two stages. First, we exploit the natural correspondence between the visual and auditory modalities in real videos to learn, in a self-supervised cross-modal manner, temporally dense video representations that capture factors such as facial movements, expression, and identity. Second, we use these learned representations as targets to be predicted by our forgery detector along with the usual binary forgery classification task; this encourages it to base its real/fake decision on said factors. We show that our method achieves state-of-the-art performance on cross-manipulation generalisation and robustness experiments, and examine the factors that contribute to its performance. Our results suggest that leveraging natural and unlabelled videos is a promising direction for the development of more robust face forgery detectors.

Rafael Poyiadzi, Jie Shen, Stavros Petridis et al.

Apparent emotional facial expression recognition has attracted a lot of research attention recently. However, the majority of approaches ignore age differences and train a generic model for all ages. In this work, we study the effect of using different age-groups for training apparent emotional facial expression recognition models. To this end, we study Domain Generalisation in the context of apparent emotional facial expression recognition from facial imagery across different age groups. We first compare several domain generalisation algorithms on the basis of out-of-domain-generalisation, and observe that the Class-Conditional Domain-Adversarial Neural Networks (CDANN) algorithm has the best performance. We then study the effect of variety and number of age-groups used during training on generalisation to unseen age-groups and observe that an increase in the number of training age-groups tends to increase the apparent emotional facial expression recognition performance on unseen age-groups. We also show that exclusion of an age-group during training tends to affect more the performance of the neighbouring age groups.

Pingchuan Ma, Rodrigo Mira, Stavros Petridis et al.

The large amount of audiovisual content being shared online today has drawn substantial attention to the prospect of audiovisual self-supervised learning. Recent works have focused on each of these modalities separately, while others have attempted to model both simultaneously in a cross-modal fashion. However, comparatively little attention has been given to leveraging one modality as a training objective to learn from the other. In this work, we propose Learning visual speech Representations from Audio via self-supervision (LiRA). Specifically, we train a ResNet+Conformer model to predict acoustic features from unlabelled visual speech. We find that this pre-trained model can be leveraged towards word-level and sentence-level lip-reading through feature extraction and fine-tuning experiments. We show that our approach significantly outperforms other self-supervised methods on the Lip Reading in the Wild (LRW) dataset and achieves state-of-the-art performance on Lip Reading Sentences 2 (LRS2) using only a fraction of the total labelled data.

Rodrigo Mira, Konstantinos Vougioukas, Pingchuan Ma et al.

Video-to-speech is the process of reconstructing the audio speech from a video of a spoken utterance. Previous approaches to this task have relied on a two-step process where an intermediate representation is inferred from the video, and is then decoded into waveform audio using a vocoder or a waveform reconstruction algorithm. In this work, we propose a new end-to-end video-to-speech model based on Generative Adversarial Networks (GANs) which translates spoken video to waveform end-to-end without using any intermediate representation or separate waveform synthesis algorithm. Our model consists of an encoder-decoder architecture that receives raw video as input and generates speech, which is then fed to a waveform critic and a power critic. The use of an adversarial loss based on these two critics enables the direct synthesis of raw audio waveform and ensures its realism. In addition, the use of our three comparative losses helps establish direct correspondence between the generated audio and the input video. We show that this model is able to reconstruct speech with remarkable realism for constrained datasets such as GRID, and that it is the first end-to-end model to produce intelligible speech for LRW (Lip Reading in the Wild), featuring hundreds of speakers recorded entirely `in the wild'. We evaluate the generated samples in two different scenarios -- seen and unseen speakers -- using four objective metrics which measure the quality and intelligibility of artificial speech. We demonstrate that the proposed approach outperforms all previous works in most metrics on GRID and LRW.

Konstantinos Vougioukas, Stavros Petridis, Maja Pantic

Domain translation is the process of transforming data from one domain to another while preserving the common semantics. Some of the most popular domain translation systems are based on conditional generative adversarial networks, which use source domain data to drive the generator and as an input to the discriminator. However, this approach does not enforce the preservation of shared semantics since the conditional input can often be ignored by the discriminator. We propose an alternative method for conditioning and present a new framework, where two networks are simultaneously trained, in a supervised manner, to perform domain translation in opposite directions. Our method is not only better at capturing the shared information between two domains but is more generic and can be applied to a broader range of problems. The proposed framework performs well even in challenging cross-modal translations, such as video-driven speech reconstruction, for which other systems struggle to maintain correspondence.

Pingchuan Ma, Stavros Petridis, Maja Pantic

In this work, we present a hybrid CTC/Attention model based on a ResNet-18 and Convolution-augmented transformer (Conformer), that can be trained in an end-to-end manner. In particular, the audio and visual encoders learn to extract features directly from raw pixels and audio waveforms, respectively, which are then fed to conformers and then fusion takes place via a Multi-Layer Perceptron (MLP). The model learns to recognise characters using a combination of CTC and an attention mechanism. We show that end-to-end training, instead of using pre-computed visual features which is common in the literature, the use of a conformer, instead of a recurrent network, and the use of a transformer-based language model, significantly improve the performance of our model. We present results on the largest publicly available datasets for sentence-level speech recognition, Lip Reading Sentences 2 (LRS2) and Lip Reading Sentences 3 (LRS3), respectively. The results show that our proposed models raise the state-of-the-art performance by a large margin in audio-only, visual-only, and audio-visual experiments.

Dominika Woszczyk, Stavros Petridis, David Millard

Speech recognition systems have improved dramatically over the last few years, however, their performance is significantly degraded for the cases of accented or impaired speech. This work explores domain adversarial neural networks (DANN) for speaker-independent speech recognition on the UAS dataset of dysarthric speech. The classification task on 10 spoken digits is performed using an end-to-end CNN taking raw audio as input. The results are compared to a speaker-adaptive (SA) model as well as speaker-dependent (SD) and multi-task learning models (MTL). The experiments conducted in this paper show that DANN achieves an absolute recognition rate of 74.91% and outperforms the baseline by 12.18%. Additionally, the DANN model achieves comparable results to the SA model's recognition rate of 77.65%. We also observe that when labelled dysarthric speech data is available DANN and MTL perform similarly, but when they are not DANN performs better than MTL.

Pingchuan Ma, Yujiang Wang, Jie Shen et al.

In this work, we present the Densely Connected Temporal Convolutional Network (DC-TCN) for lip-reading of isolated words. Although Temporal Convolutional Networks (TCN) have recently demonstrated great potential in many vision tasks, its receptive fields are not dense enough to model the complex temporal dynamics in lip-reading scenarios. To address this problem, we introduce dense connections into the network to capture more robust temporal features. Moreover, our approach utilises the Squeeze-and-Excitation block, a light-weight attention mechanism, to further enhance the model's classification power. Without bells and whistles, our DC-TCN method has achieved 88.36% accuracy on the Lip Reading in the Wild (LRW) dataset and 43.65% on the LRW-1000 dataset, which has surpassed all the baseline methods and is the new state-of-the-art on both datasets.

Pingchuan Ma, Brais Martinez, Stavros Petridis et al.

Lipreading has witnessed a lot of progress due to the resurgence of neural networks. Recent works have placed emphasis on aspects such as improving performance by finding the optimal architecture or improving generalization. However, there is still a significant gap between the current methodologies and the requirements for an effective deployment of lipreading in practical scenarios. In this work, we propose a series of innovations that significantly bridge that gap: first, we raise the state-of-the-art performance by a wide margin on LRW and LRW-1000 to 88.5% and 46.6%, respectively using self-distillation. Secondly, we propose a series of architectural changes, including a novel Depthwise Separable Temporal Convolutional Network (DS-TCN) head, that slashes the computational cost to a fraction of the (already quite efficient) original model. Thirdly, we show that knowledge distillation is a very effective tool for recovering performance of the lightweight models. This results in a range of models with different accuracy-efficiency trade-offs. However, our most promising lightweight models are on par with the current state-of-the-art while showing a reduction of 8.2x and 3.9x in terms of computational cost and number of parameters, respectively, which we hope will enable the deployment of lipreading models in practical applications.

Abhinav Shukla, Stavros Petridis, Maja Pantic

The intuitive interaction between the audio and visual modalities is valuable for cross-modal self-supervised learning. This concept has been demonstrated for generic audiovisual tasks like video action recognition and acoustic scene classification. However, self-supervision remains under-explored for audiovisual speech. We propose a method to learn self-supervised speech representations from the raw audio waveform. We train a raw audio encoder by combining audio-only self-supervision (by predicting informative audio attributes) with visual self-supervision (by generating talking faces from audio). The visual pretext task drives the audio representations to capture information related to lip movements. This enriches the audio encoder with visual information and the encoder can be used for evaluation without the visual modality. Our method attains competitive performance with respect to existing self-supervised audio features on established isolated word classification benchmarks, and significantly outperforms other methods at learning from fewer labels. Notably, our method also outperforms fully supervised training, thus providing a strong initialization for speech related tasks. Our results demonstrate the potential of multimodal self-supervision in audiovisual speech for learning good audio representations.

Abhinav Shukla, Stavros Petridis, Maja Pantic

Self-supervised learning has attracted plenty of recent research interest. However, most works for self-supervision in speech are typically unimodal and there has been limited work that studies the interaction between audio and visual modalities for cross-modal self-supervision. This work (1) investigates visual self-supervision via face reconstruction to guide the learning of audio representations; (2) proposes an audio-only self-supervision approach for speech representation learning; (3) shows that a multi-task combination of the proposed visual and audio self-supervision is beneficial for learning richer features that are more robust in noisy conditions; (4) shows that self-supervised pretraining can outperform fully supervised training and is especially useful to prevent overfitting on smaller sized datasets. We evaluate our learned audio representations for discrete emotion recognition, continuous affect recognition and automatic speech recognition. We outperform existing self-supervised methods for all tested downstream tasks. Our results demonstrate the potential of visual self-supervision for audio feature learning and suggest that joint visual and audio self-supervision leads to more informative audio representations for speech and emotion recognition.

Brais Martinez, Pingchuan Ma, Stavros Petridis et al.

Lip-reading has attracted a lot of research attention lately thanks to advances in deep learning. The current state-of-the-art model for recognition of isolated words in-the-wild consists of a residual network and Bidirectional Gated Recurrent Unit (BGRU) layers. In this work, we address the limitations of this model and we propose changes which further improve its performance. Firstly, the BGRU layers are replaced with Temporal Convolutional Networks (TCN). Secondly, we greatly simplify the training procedure, which allows us to train the model in one single stage. Thirdly, we show that the current state-of-the-art methodology produces models that do not generalize well to variations on the sequence length, and we addresses this issue by proposing a variable-length augmentation. We present results on the largest publicly-available datasets for isolated word recognition in English and Mandarin, LRW and LRW1000, respectively. Our proposed model results in an absolute improvement of 1.2% and 3.2%, respectively, in these datasets which is the new state-of-the-art performance.