Audio comprehension-including speech, non-speech sounds, and music-is essential for achieving human-level intelligence. Consequently, AI agents must demonstrate holistic audio understanding to qualify as generally intelligent. However, evaluating auditory intelligence comprehensively remains challenging. To address this gap, we introduce MMAU-Pro, the most comprehensive and rigorously curated benchmark for assessing audio intelligence in AI systems. MMAU-Pro contains 5,305 instances, where each instance has one or more audios paired with human expert-generated question-answer pairs, spanning speech, sound, music, and their combinations. Unlike existing benchmarks, MMAU-Pro evaluates auditory intelligence across 49 unique skills and multiple complex dimensions, including long-form audio comprehension, spatial audio reasoning, multi-audio understanding, among others. All questions are meticulously designed to require deliberate multi-hop reasoning, including both multiple-choice and open-ended response formats. Importantly, audio data is sourced directly ``from the wild" rather than from existing datasets with known distributions. We evaluate 22 leading open-source and proprietary multimodal AI models, revealing significant limitations: even state-of-the-art models such as Gemini 2.5 Flash and Audio Flamingo 3 achieve only 59.2% and 51.7% accuracy, respectively, approaching random performance in multiple categories. Our extensive analysis highlights specific shortcomings and provides novel insights, offering actionable perspectives for the community to enhance future AI systems' progression toward audio general intelligence. The benchmark and code is available at https://sonalkum.github.io/mmau-pro.

Abstract Large-scale vision-language models demonstrate strong multimodal alignment and generalization across diverse tasks. Among them, CLIP stands out as one of the most successful approaches. In this work, we extend the application of CLIP to sound source localization, proposing a self-supervised method operates without explicit text input. We introduce a framework that maps audios into tokens compatible with CLIP’s text encoder, producing audio-driven embeddings. These embeddings are used to generate sounding region masks, from which visual features are extracted and aligned with the audio embeddings through a contrastive audio-visual correspondence objective. Our findings show that alignment knowledge of pre-trained multimodal foundation model enables our method to generate more complete and compact localization for sounding objects. We further propose an LLM-guided extension that distills object-aware audio-visual scene understanding into the model during training to enhance alignment. Extensive experiments across five diverse tasks demonstrate that our method, in all variants, outperforms state-of-the-art approaches and achieves strong generalization in zero-shot settings.

The goal of this letter is to enhance Text-to-Audio (T2A) generation at inference, focusing on generating realistic audio that precisely aligns with text prompts. Despite the rapid advancements, existing approaches often fail to achieve a reliable balance between semantic alignment and perceptual quality. We address this by adopting Inference-Time Scaling for the first time in T2A generation, and propose a multi-reward guidance that places each criterion on equal footing. By standardizing each reward to zero mean and unit variance before a weighted summation, the method reduces the scale mismatch that otherwise demands weight tuning, providing stable guidance at equal weights while leaving the weight as an interpretable control over the desired aspect. Moreover, we introduce a new text-audio alignment metric using an audio language model for more robust evaluation. Empirically, our method improves both semantic alignment and perceptual quality, achieving improved performance over naive generation and existing reward guidance techniques.

The objective of this study is to generate high-quality speech from silent talking face videos, a task also known as video-to-speech synthesis. A significant challenge in video-to-speech synthesis lies in the substantial modality gap between silent video and multi-faceted speech. In this paper, we propose a novel video-to-speech system that effectively bridges this modality gap, significantly enhancing the quality of synthesized speech. This is achieved by learning of hierarchical representations from video to speech. Specifically, we gradually transform silent video into acoustic feature spaces through three sequential stages – content, timbre, and prosody modeling. In each stage, we align visual factors – lip movements, face identity, and facial expressions – with corresponding acoustic counterparts to ensure the seamless transformation. Additionally, to generate realistic and coherent speech from the visual representations, we employ a flow matching model that estimates direct trajectories from a simple prior distribution to the target speech distribution. Extensive experiments demonstrate that our method achieves exceptional generation quality comparable to real utterances, outperforming existing methods by a significant margin.

We present a unified model capable of simultaneously grounding both spoken language and non-speech sounds within a visual scene, addressing key limitations in current audio-visual grounding models. Existing approaches are typically limited to handling either speech or non-speech sounds independently, or at best, together but sequentially without mixing. This limitation prevents them from capturing the complexity of real-world audio sources that are often mixed. Our approach introduces a "mix-and-separate" framework with audio-visual alignment objectives that jointly learn correspondence and disentanglement using mixed audio. Through these objectives, our model learns to produce distinct embeddings for each audio type, enabling effective disentanglement and grounding across mixed audio sources. Additionally, we created a new dataset to evaluate simultaneous grounding of mixed audio sources, demonstrating that our model outperforms prior methods. Our approach also achieves comparable or better performance in standard segmentation and cross-modal retrieval tasks, highlighting the benefits of our mix-and-separate approach.

The goal of this paper is to enhance face recognition performance by augmenting head poses during the testing phase. Existing methods often rely on training on frontalised images or learning pose-invariant representations, yet both approaches typically require re-training and testing for each dataset, involving a substantial amount of effort. In contrast, this study proposes Pose-TTA, a novel approach that aligns faces at inference time without additional training. To achieve this, we employ a portrait animator that transfers the source image identity into the pose of a driving image. Instead of frontalising a side-profile face - which can introduce distortion - Pose-TTA generates matching side-profile images for comparison, thereby reducing identity information loss. Furthermore, we propose a weighted feature aggregation strategy to address any distortions or biases arising from the synthetic data, thus enhancing the reliability of the augmented images. Extensive experiments on diverse datasets and with various pre-trained face recognition models demonstrate that PoseTTA consistently improves inference performance. Moreover, our method is straightforward to integrate into existing face recognition pipelines, as it requires no retraining or fine-tuning of the underlying recognition models.

In this paper, we introduce V2SFlow, a novel Video-to-Speech (V2S) framework designed to generate natural and intelligible speech directly from silent talking face videos. While recent V2S systems have shown promising results on constrained datasets with limited speakers and vocabularies, their performance often degrades on real-world, unconstrained datasets due to the inherent variability and complexity of speech signals. To address these challenges, we decompose the speech signal into manageable subspaces (content, pitch, and speaker information), each representing distinct speech attributes, and predict them directly from the visual input. To generate coherent and realistic speech from these predicted attributes, we employ a rectified flow matching decoder built on a Transformer architecture, which models efficient probabilistic pathways from random noise to the target speech distribution. Extensive experiments demonstrate that V2SFlow significantly outperforms state-of-the-art methods, even surpassing the naturalness of ground truth utterances.

We present VoiceDiT, a multi-modal generative model for producing environment-aware speech and audio from text and visual prompts. While aligning speech with text is crucial for intelligible speech, achieving this alignment in noisy conditions remains a significant and underexplored challenge in the field. To address this, we present a novel audio generation pipeline named VoiceDiT. This pipeline includes three key components: (1) the creation of a large-scale synthetic speech dataset for pre-training and a refined real-world speech dataset for fine-tuning, (2) the Dual-DiT, a model designed to efficiently preserve aligned speech information while accurately reflecting environmental conditions, and (3) a diffusion-based Image-to-Audio Translator that allows the model to bridge the gap between audio and image, facilitating the generation of environmental sound that aligns with the multi-modal prompts. Extensive experimental results demonstrate that VoiceDiT outperforms previous models on real-world datasets, showcasing significant improvements in both audio quality and modality integration. Synthesized samples are available on our demo page: https://mm.kaist.ac.kr/projects/voicedit/

The goal of voice conversion is to transform the speech of a source speaker to sound like that of a reference speaker while preserving the original content. A key challenge is to extract disentangled linguistic content from the source and voice style from the reference. While existing approaches leverage various methods to isolate the two, a generalization still requires further attention, especially for robustness in zero-shot scenarios. In this paper, we achieve successful disentanglement of content and speaker features by tuning self-supervised speech features with adapters. The adapters are trained to dynamically encode nuanced features from rich self-supervised features, and the decoder fuses them to produce speech that accurately resembles the reference with minimal loss of content. Moreover, we leverage a conditional flow matching decoder with cross-attention speaker conditioning to further boost the synthesis quality and efficiency. Subjective and objective evaluations in a zero-shot scenario demonstrate that the proposed method outperforms existing models in speech quality and similarity to the reference speech.

The goal of this paper is to accelerate codec-based speech synthesis systems with minimum sacrifice to speech quality. We propose an enhanced inference method that allows for flexible trade-offs between speed and quality during inference without requiring additional training. Our core idea is to predict multiple tokens per inference step of the AR module using multiple prediction heads, resulting in a linear reduction in synthesis time as the number of heads increases. Furthermore, we introduce a novel speculative decoding technique that utilises a Viterbi-based algorithm to select the optimal sequence of generated tokens at each decoding step. In our experiments, we demonstrate that the time required to predict each token is reduced by a factor of 4 to 5 compared to baseline models, with minimal quality trade-off or even improvement in terms of speech intelligibility. Audio samples are available at: mm.kaist.ac.kr/projects/mtp.

Automated audio captioning is a task that generates textual descriptions for audio content, and recent studies have explored using visual information to enhance captioning quality. However, current methods often fail to effectively fuse audio and visual data, missing important semantic cues from each modality. To address this, we introduce LAVCap, a large language model (LLM)-based audio-visual captioning framework that effectively integrates visual information with audio to improve audio captioning performance. LAVCap employs an optimal transport-based alignment loss to bridge the modality gap between audio and visual features, enabling more effective semantic extraction. Additionally, we propose an optimal transport attention module that enhances audio-visual fusion using an optimal transport assignment map. Combined with the optimal training strategy, experimental results demonstrate that each component of our framework is effective. LAVCap outperforms existing state-of-the-art methods on the AudioCaps dataset, without relying on large datasets or post-processing. Code is available at https://github.com/NAVER-INTEL-Co-Lab/gaudi-lavcap.

We present VoiceCraft-Dub, a novel approach for automated video dubbing that synthesizes high-quality speech from text and facial cues. This task has broad applications in filmmaking, multimedia creation, and assisting voice-impaired individuals. Building on the success of Neural Codec Language Models (NCLMs) for speech synthesis, our method extends their capabilities by incorporating video features, ensuring that synthesized speech is time-synchronized and expressively aligned with facial movements while preserving natural prosody. To inject visual cues, we design adapters to align facial features with the NCLM token space and introduce audio-visual fusion layers to merge audio-visual information within the NCLM framework. Additionally, we curate CelebV-Dub, a new dataset of expressive, real-world videos specifically designed for automated video dubbing. Extensive experiments show that our model achieves high-quality, intelligible, and natural speech synthesis with accurate lip synchronization, outperforming existing methods in human perception and performing favorably in objective evaluations. We also adapt VoiceCraft-Dub for the video-to-speech task, demonstrating its versatility for various applications.

Generative modeling has seen significant advancements in image and video synthesis. However, the curse of dimensionality remains a significant obstacle, especially for video generation, given its inherently complex and high-dimensional nature. Many existing works rely on low-dimensional latent spaces from pretrained image autoencoders. However, this approach overlooks temporal redundancy in videos and often leads to temporally incoherent decoding. To address this issue, we propose a video compression network that reduces the dimensionality of visual data both spatially and temporally. Our model, based on a variational autoencoder, employs causal 3D convolution to handle images and videos jointly. The key contributions of our work include a scale-agnostic encoder for preserving video fidelity, a novel spatio-temporal down/upsampling block for robust long-sequence modeling, and a flow regularization loss for accurate motion decoding. Our approach outperforms competitors in video quality and compression rates across various datasets. Experimental analyses also highlight its potential as a robust autoencoder for video generation training.

Following the success of Large Language Models (LLMs), expanding their boundaries to new modalities represents a significant paradigm shift in multimodal understanding. Human perception is inherently multimodal, relying not only on text but also on auditory and visual cues for a complete understanding of the world. In recognition of this fact, audio-visual LLMs have recently emerged. Despite promising developments, the lack of dedicated benchmarks poses challenges for understanding and evaluating models. In this work, we show that audio-visual LLMs struggle to discern subtle relationships between audio and visual signals, leading to hallucinations and highlighting the need for reliable benchmarks. To address this, we introduce AVHBench, the first comprehensive benchmark specifically designed to evaluate the perception and comprehension capabilities of audio-visual LLMs. Our benchmark includes tests for assessing hallucinations, as well as the cross-modal matching and reasoning abilities of these models. Our results reveal that most existing audio-visual LLMs struggle with hallucinations caused by cross-interactions between modalities, due to their limited capacity to perceive complex multimodal signals and their relationships. Additionally, we demonstrate that simple training with our AVHBench improves robustness of audio-visual LLMs against hallucinations. Dataset: https://github.com/kaist-ami/AVHBench

In this paper, we address the task of multimodal-to-speech generation, which aims to synthesize high-quality speech from multiple input modalities: text, video, and reference audio. This task has gained increasing attention due to its wide range of applications, such as film production, dubbing, and virtual avatars. Despite recent progress, existing methods still suffer from limitations in speech intelligibility, audio-video synchronization, speech naturalness, and voice similarity to the reference speaker. To address these challenges, we propose AlignDiT, a multimodal Aligned Diffusion Transformer that generates accurate, synchronized, and natural-sounding speech from aligned multimodal inputs. Built upon the in-context learning capability of the DiT architecture, AlignDiT explores three effective strategies to align multimodal representations. Furthermore, we introduce a novel multimodal classifier-free guidance mechanism that allows the model to adaptively balance information from each modality during speech synthesis. Extensive experiments demonstrate that AlignDiT significantly outperforms existing methods across multiple benchmarks in terms of quality, synchronization, and speaker similarity. Moreover, AlignDiT exhibits strong generalization capability across various multimodal tasks, such as video-to-speech synthesis and visual forced alignment, consistently achieving state-of-the-art performance. The demo page is available at https://mm.kaist.ac.kr/projects/AlignDiT.

Hallucination remains a major challenge in multimodal large language models (MLLMs). To address this, various contrastive decoding (CD) methods have been proposed that contrasts original logits with hallucinated logits generated from perturbed inputs. While CD has shown promise in vision-language models (VLMs), it is not well-suited for AV-LLMs, where hallucinations often emerge from both unimodal and cross-modal combinations involving audio, video, and language. These intricate interactions call for a more adaptive and modality-aware decoding strategy. In this paper, we propose Audio-Visual Contrastive Decoding (AVCD)—a novel, training-free decoding framework designed to model trimodal interactions and suppress modality-induced hallucinations in AV-LLMs. Unlike previous CD methods in VLMs that corrupt a fixed modality, AVCD leverages attention distributions to dynamically identify less dominant modalities and applies attentive masking to generate perturbed output logits. To support CD in a trimodal setting, we also reformulate the original CD framework to jointly handle audio, visual, and textual inputs. Finally, to improve efficiency, we introduce entropy-guided adaptive decoding, which selectively skips unnecessary decoding steps based on the model’s confidence in its predictions. Extensive experiments demonstrate that AVCD consistently outperforms existing decoding methods. Especially, on the AVHBench dataset, it improves accuracy by 2% for VideoLLaMA2 and 7% for video-SALMONN, demonstrating strong robustness and generalizability. Our code is available at : https://github.com/kaistmm/AVCD.

We present MGAudio, a novel flow-based framework for open-domain video-to-audio generation, which introduces model-guided dual-role alignment as a central design principle. Unlike prior approaches that rely on classifier-based or classifier-free guidance, MGAudio enables the generative model to guide itself through a dedicated training objective designed for video-conditioned audio generation. The framework integrates three main components: (1) a scalable flow-based Transformer denoiser, (2) a dual-role alignment mechanism where the audio-visual encoder serves both as a conditioning module and as a feature aligner to improve generation quality, and (3) a model-guided objective that enhances cross-modal coherence and audio realism. MGAudio achieves state-of-the-art performance on VGGSound, reducing FAD to 0.40, substantially surpassing the best classifier-free guidance baselines, and consistently outperforms existing methods across FD, IS, and alignment metrics. It also generalizes well to the challenging UnAV-100 benchmark. These results highlight model-guided dual-role alignment as a powerful and scalable paradigm for conditional video-to-audio generation. Code is available at: https://github.com/pantheon5100/mgaudio

Distinguishing visually similar objects by their motion remains a critical challenge in computer vision. Although supervised trackers show promise, contemporary self-supervised trackers struggle when visual cues become ambiguous, limiting their scalability and generalization without extensive labeled data. We find that pre-trained video diffusion models inherently learn motion representations suitable for tracking without task-specific training. This ability arises because their denoising process isolates motion in early, high-noise stages, distinct from later appearance refinement. Capitalizing on this discovery, our self-supervised tracker significantly improves performance in distinguishing visually similar objects, an underexplored failure point for existing methods. Our method achieves up to a 6-point improvement over recent self-supervised approaches on established benchmarks and our newly introduced tests focused on tracking visually similar items. Visualizations confirm that these diffusion-derived motion representations enable robust tracking of even identical objects across challenging viewpoint changes and deformations. Project page: \small{\url{https://chenshuang-zhang.github.io/projects/ted}}.

Recent studies on learning-based sound source localization have primarily focused on localization performance. However, prior work and existing benchmarks often overlook a crucial aspect: cross-modal interaction, which is essential for interactive sound source localization. This interaction is vital for understanding semantically matched or mismatched audio-visual events, such as silent objects or true sound sources among multiple objects. In this work, we comprehensively examine the cross-modal interaction of existing methods, benchmarks, evaluation metrics, and cross-modal understanding tasks. We identify the overlooked points of previous studies and make several contributions to address them. First, we propose a learning framework that incorporates retrieval-based and hand-crafted augmentation techniques, enhancing cross-modal interaction through cross-modal alignment. Second, we introduce new evaluation metrics to accurately and rigorously assess localization methods, focusing on both localization performance and cross-modal interaction. Third, to thoroughly analyze interactive sound source localization, we present a new semi-synthetic benchmark with diverse categorical combinations. Finally, we evaluate both interactive sound source localization and auxiliary cross-modal retrieval tasks, benchmarking competing methods alongside our own. Our new benchmark and evaluation metrics reveal that previous methods struggle with interactive sound source localization tasks, largely due to their limited cross-modal interaction capabilities. Our method, which features enhanced cross-modal alignment, demonstrates superior sound source localization and cross-modal interaction performance. This work provides the most comprehensive analysis of sound source localization to date, with extensive validation of competing methods on both existing and new benchmarks using both new and standard evaluation metrics.

The goal of this work is to reconstruct high quality speech from lip motions alone, a task also known as lip-to-speech. A key challenge of lip-to-speech systems is the one-to-many mapping caused by (1) the existence of homophenes and (2) multiple speech variations, resulting in a mispronounced and over-smoothed speech. In this paper, we propose a novel lip-to-speech system that significantly improves the generation quality by alleviating the one-to-many mapping problem from multiple perspectives. Specifically, we incorporate (1) self-supervised speech representations to disambiguate homophenes, and (2) acoustic variance information to model diverse speech styles. Additionally, to better solve the aforementioned problem, we employ a flow based post-net which captures and refines the details of the generated speech. We perform extensive experiments on two datasets, and demonstrate that our method achieves the generation quality close to that of real human utterance, outperforming existing methods in terms of speech naturalness and intelligibility by a large margin. Synthesised samples are available at our demo page: https://mm.kaist.ac.kr/projects/LTBS.

Long-form video content constitutes a significant portion of internet traffic, making automated video summarization an essential research problem. However, existing video summarization datasets are notably limited in their size, constraining the effectiveness of state-of-the-art methods for generalization. Our work aims to overcome this limitation by capitalizing on the abundance of long-form videos with dense speech-to-video alignment and the remarkable capabilities of recent large language models (LLMs) in summarizing long text. We introduce an automated and scalable pipeline for generating a large-scale video summarization dataset using LLMs as Oracle summarizers. By leveraging the generated dataset, we analyze the limitations of existing approaches and propose a new video summarization model that effectively addresses them. To facilitate further research in the field, our work also presents a new benchmark dataset that contains 1200 long videos each with high-quality summaries annotated by professionals. Extensive experiments clearly indicate that our proposed approach sets a new state-of-the-art in video summarization across several benchmarks.

Movie trailers are an essential tool for promoting films and attracting audiences. However, the process of creating trailers can be time-consuming and expensive. To stream-line this process, we propose an automatic trailer generation framework that generates plausible trailers from a full movie by automating shot selection and composition. Our approach draws inspiration from machine translation techniques and models the movies and trailers as sequences of shots, thus formulating the trailer generation problem as a sequence-to-sequence task. We introduce Trailer Generation Transformer (TGT), a deep-learning framework utilizing an encoder-decoder architecture. TGT movie encoder is tasked with contextualizing each movie shot representation via self-attention, while the autoregressive trailer de-coder predicts the feature representation of the next trailer shot, accounting for the relevance of shots' temporal order in trailers. Our TGT significantly outperforms previous methods on a comprehensive suite of metrics.

The goal of this work is to simultaneously generate natural talking faces and speech outputs from text. We achieve this by integrating Talking Face Generation (TFG) and Text-to-Speech (TTS) systems into a unified framework. We address the main challenges of each task: (1) generating a range of head poses representative of real-world scenarios, and (2) ensuring voice consistency despite variations in facial motion for the same identity. To tackle these issues, we introduce a motion sampler based on conditional flow matching, which is capable of high-quality motion code generation in an efficient way. Moreover, we introduce a novel conditioning method for the TTS system, which utilises motion-removed features from the TFG model to yield uniform speech outputs. Our extensive experiments demonstrate that our method effectively creates natural-looking talking faces and speech that accurately match the input text. To our knowledge, this is the first effort to build a multimodal synthesis system that can generalise to unseen identities.

The objective of this work is the effective extraction of spatial and dynamic features for Continuous Sign Language Recognition (CSLR). To accomplish this, we utilise a two-pathway SlowFast network, where each pathway operates at distinct temporal resolutions to separately capture spatial (hand shapes, facial expressions) and dynamic (movements) information. In addition, we introduce two distinct feature fusion methods, carefully designed for the characteristics of CSLR: (1) Bi-directional Feature Fusion (BFF), which facilitates the transfer of dynamic semantics into spatial semantics and vice versa; and (2) Pathway Feature Enhancement (PFE), which enriches dynamic and spatial representations through auxiliary subnetworks, while avoiding the need for extra inference time. As a result, our model further strengthens spatial and dynamic representations in parallel. We demonstrate that the proposed framework outperforms the current state-of-the-art performance on popular CSLR datasets, including PHOENIX14, PHOENIX14-T, and CSL-Daily.

Transformers have become central to recent advances in audio classification. However, training an audio spectrogram transformer, e.g. AST, from scratch can be resource and time-intensive. Furthermore, the complexity of transformers heavily depends on the input audio spectrogram size. In this work, we aim to optimize AST training by linking to the resolution in the time-axis. We introduce multi-phase training of audio spectrogram transformers by connecting the seminal idea of coarse-to-fine with transformer models. To achieve this, we propose a set of methods for temporal compression. By employing one of these methods, the transformer model learns from lowerresolution (coarse) data in the initial phases, and then is fine-tuned with high-resolution data later in a curriculum learning strategy. Experimental results demonstrate that the proposed training mechanism for AST leads to improved (or on-par) performance with faster convergence, i.e. requiring fewer computational resources and less time. This approach is also generalizable to other AST-based methods regardless of their learning paradigms.

In the field of speaker verification, session or channel variability poses a significant challenge. While many contemporary methods aim to disentangle session information from speaker embeddings, we introduce a novel approach using an additional embedding to represent the session information. This is achieved by training an auxiliary network appended to the speaker embedding extractor which remains fixed in this training process. This results in two similarity scores: one for the speakers information and one for the session information. The latter score acts as a compensator for the former that might be skewed due to session variations. Our extensive experiments demonstrate that session information can be effectively compensated without retraining of the embedding extractor.

The goal of this work is Active Speaker Detection (ASD), a task to determine whether a person is speaking or not in a series of video frames. Previous works have dealt with the task by exploring network architectures while learning effective representations has been less explored. In this work, we propose TalkNCE, a novel talk-aware contrastive loss. The loss is only applied to part of the full segments where a person on the screen is actually speaking. This encourages the model to learn effective representations through the natural correspondence of speech and facial movements. Our loss can be jointly optimized with the existing objectives for training ASD models without the need for additional supervision or training data. The experiments demonstrate that our loss can be easily integrated into the existing ASD frameworks, improving their performance. Our method achieves state-of-the-art performances on AVA-ActiveSpeaker and ASW datasets.

This paper presents a multi-modal dataset that contains rich transcriptions of spoken conversations. As diverse multi-modal and multi-task models emerge, there is a growing need for multi-modal training and evaluation datasets accompanied by rich metadata. However, there is no universal dataset that addresses these requirements for the diverse tasks partially due to the cost of annotation. To overcome this limitation, we develop a semi-automatic pipeline that makes the annotation more feasible. The resulting dataset is VoxMM, a multi-modal, multi-domain dataset. VoxMM incorporates video, audio, and text modalities. In terms of labels, it offers a wide array of metadata such as speaker labels, transcriptions, gender, and more. VoxMM supports both the training and the evaluation of any-to-any modality mapping models. It also offers a more accurate representation of real-world scenarios, bridging the gap between controlled laboratory experiments and the varying performances in the real-world. We present initial benchmarks on automatic speech recognition and speaker diarisation. The VoxMM dataset can be downloaded from https://mm.kaist.ac.kr/projects/voxmm

The objective of this work is to extract the target speaker’s voice from a mixture of voices using visual cues. Existing works on audio-visual speech separation have demonstrated their performance with promising intelligibility, but maintaining naturalness remains challenging. To address this issue, we propose AVDiffuSS, an audio-visual speech separation model based on a diffusion mechanism known for its capability to generate natural samples. We also propose a cross-attention-based feature fusion mechanism for an effective fusion of the two modalities for diffusion. This mechanism is specifically tailored for the speech domain to integrate the phonetic information from audio-visual correspondence in speech generation. In this way, the fusion process maintains the high temporal resolution of the features, without excessive computational requirements. We demonstrate that the proposed framework achieves state-of-the-art results on two benchmarks, including VoxCeleb2 and LRS3, producing speech with notably better naturalness. Project page with demo: https://mm.kaist.ac.kr/projects/avdiffuss/

This paper presents VoiceLDM, a model designed to produce audio that accurately follows two distinct natural language text prompts: the description prompt and the content prompt. The former provides information about the overall environmental context of the audio, while the latter conveys the linguistic content. To achieve this, we adopt a text-to-audio (TTA) model based on latent diffusion models and extend its functionality to incorporate an additional content prompt as a conditional input. By utilizing pretrained contrastive language-audio pretraining (CLAP) and Whisper, VoiceLDM is trained on large amounts of real-world audio without manual annotations or transcriptions. Additionally, we employ dual classifier-free guidance to further enhance the controllability of VoiceLDM. Experimental results demonstrate that VoiceLDM is capable of generating plausible audio that aligns well with both input conditions, even surpassing the speech intelligibility of the ground truth audio on the AudioCaps test set. Furthermore, we explore the text-to-speech (TTS) and zero-shot text-to-audio capabilities of VoiceLDM and show that it achieves competitive results. Demos and code are available at https://voiceldm.github.io.

The goal of this paper is to generate realistic audio with a lightweight and fast diffusion-based vocoder, named FreGrad. Our framework consists of the following three key components: (1) We employ discrete wavelet transform that decomposes a complicated waveform into sub-band wavelets, which helps FreGrad to operate on a simple and concise feature space, (2) We design a frequency-aware dilated convolution that elevates frequency awareness, resulting in generating speech with accurate frequency information, and (3) We introduce a bag of tricks that boosts the generation quality of the proposed model. In our experiments, FreGrad achieves 3.7 times faster training time and 2.2 times faster inference speed compared to our baseline while reducing the model size by 0.6 times (only 1.78M parameters) without sacrificing the output quality. Audio samples are available at: https://mm.kaist.ac.kr/projects/FreGrad.

The objective of this study is to investigate the learning process of Visually Grounded Speech (VGS) models through joint learning that combines contrastive learning and masked image modeling. Typically, VGS models ahn to establish audio-visual alignment between images and then spoken captions within a contrastive learning framework. Building upon this seminal concept, in this work, we explore whether visual reconstruction with the help of cross-modality can enhance alignment, given that spoken captions describe visual appearances. To achieve this, we extend the contrastive learning-based VGS models by incorporating a masked autoencoder that utilizes cross-attention in the decoder. Through this cross-modal interaction in the decoder, spoken caption features guide the model to reconstruct the masked patches and capture correspondence between the two modalities. Our findings suggest that integrating cross-modal reconstruction within the contrastive learning framework enhances audio-visual feature alignment. Consequently, our proposed method gives comparable performance to existing models that utilize prior knowledge or other modalities, such as object region proposals or Contrastive Language-Image Pretraining (CLIP).

Recent advancements in self-supervised audio-visual representation learning have demonstrated its potential to capture rich and comprehensive representations. However, despite the advantages of data augmentation verified in many learning methods, audio-visual learning has struggled to fully harness these benefits, as augmentations can easily disrupt the correspondence between input pairs. To address this limitation, we introduce EquiAV, a novel framework that leverages equivariance for audio-visual contrastive learning. Our approach begins with extending equivariance to audio-visual learning, facilitated by a shared attention-based transformation predictor. It enables the aggregation of features from diverse augmentations into a representative embedding, providing robust supervision. Notably, this is achieved with minimal computational overhead. Extensive ablation studies and qualitative results verify the effectiveness of our method. EquiAV outperforms previous works across various audio-visual benchmarks. The code is available on https://github.com/JongSuk1/EquiAV

Video Temporal Grounding (VTG) aims to identify visual frames in a video clip that match text queries. Recent studies in VTG employ cross-attention to correlate visual frames and text queries as individual token sequences. However, these approaches overlook a crucial aspect of the problem: a holistic understanding of the query sentence. A model may capture correlations between individual word tokens and arbitrary visual frames while possibly missing out on the global meaning. To address this, we introduce two primary contributions: (1) a visual frame-level gate mechanism that incorporates holistic textual information, (2) cross-modal alignment loss to learn the fine-grained correlation between query and relevant frames. As a result, we regularize the effect of individual word tokens and suppress irrelevant visual frames. We demonstrate that our method outperforms state-of-the-art approaches in VTG benchmarks, indicating that holistic text understanding guides the model to focus on the semantically important parts within the video.

Large-scale pre-trained image-text models demonstrate remarkable versatility across diverse tasks, benefiting from their robust representational capabilities and effective multimodal alignment. We extend the application of these models, specifically CLIP, to the domain of sound source localization. Unlike conventional approaches, we employ the pre-trained CLIP model without explicit text input, relying solely on the audio-visual correspondence. To this end, we introduce a framework that translates audio signals into tokens compatible with CLIP’s text encoder, yielding audio-driven embeddings. By directly using these embeddings, our method generates audio-grounded masks for the provided audio, extracts audio-grounded image features from the highlighted regions, and aligns them with the audio-driven embeddings using the audio-visual correspondence objective. Our findings suggest that utilizing pre-trained image-text models enable our model to generate more complete and compact localization maps for the sounding objects. Extensive experiments show that our method outperforms state-of-the-art approaches by a significant margin.

Transformers have rapidly become the preferred choice for audio classification, surpassing methods based on CNNs. However, Audio Spectrogram Transformers (ASTs) exhibit quadratic scaling due to self-attention. The removal of this quadratic self-attention cost presents an appealing direction. Recently, state space models (SSMs), such as Mamba, have demonstrated potential in language and vision tasks in this regard. In this study, we explore whether reliance on self-attention is necessary for audio classification tasks. By introducing Audio Mamba (AuM), the first self-attention-free, purely SSM-based model for audio classification, we aim to address this question. We evaluate AuM on various audio datasets - comprising six different benchmarks - where it achieves comparable or better performance compared to well-established AST model.

This letter proposes a novel user-defined keyword spotting framework that accurately detects audio keywords based on text enrollment. Since audio data possesses additional acoustic information compared to text, there are discrepancies between these two modalities. To address this challenge, we present ParallelKWS, which utilises self- and cross-attention in a parallel architecture to effectively capture information both within and across the two modalities. We further propose a phoneme duration-based alignment loss that enforces the sequential correspondence between audio and text features. Extensive experimental results demonstrate that our proposed method achieves state-of-the-art performance on several benchmark datasets in both seen and unseen domains, without incorporating extra data beyond the dataset used in previous studies.

The VoxCeleb Speaker Recognition Challenges (VoxSRC) were a series of challenges and workshops that ran annually from 2019 to 2023. The challenges primarily evaluated the tasks of speaker recognition and diarisation under various settings including: closed and open training data; as well as supervised, self-supervised, and semi-supervised training for domain adaptation. The challenges also provided publicly available training and evaluation datasets for each task and setting, with new test sets released each year. In this paper, we provide a review of these challenges that covers: what they explored; the methods developed by the challenge participants and how these evolved; and also the current state of the field for speaker verification and diarisation. We chart the progress in performance over the five installments of the challenge on a common evaluation dataset and provide a detailed analysis of how each year's special focus affected participants' performance. This paper is aimed both at researchers who want an overview of the speaker recognition and diarisation field, and also at challenge organisers who want to benefit from the successes and avoid the mistakes of the VoxSRC challenges. We end with a discussion of the current strengths of the field and open challenges.

The goal of this work is to develop self-sufficient framework for Continuous Sign Language Recognition (CSLR) that addresses key issues of sign language recognition. These include the need for complex multi-scale features such as hands, face, and mouth for understanding, and absence of frame-level annotations. To this end, we propose (1) Divide and Focus Convolution (DFConv) which extracts both manual and non-manual features without the need for additional networks or annotations, and (2) Dense Pseudo-Label Refinement (DPLR) which propagates non-spiky frame-level pseudo-labels by combining the ground truth gloss sequence labels with the predicted sequence. We demonstrate that our model achieves state-of-the-art performance among RGB-based methods on large-scale CSLR benchmarks, PHOENIX-2014 and PHOENIX-2014-T, while showing comparable results with better efficiency when compared to other approaches that use multi-modality or extra annotations.

Speaker embedding extractors (EEs), which map input audio to a speaker discriminant latent space, are of paramount importance in speaker diarisation. However, there are several challenges when adopting EEs for diarisation, from which we tackle two key problems. First, the evaluation is not straightforward because the required features differ between speaker verification and diarisation. We show that better performance on widely adopted speaker verification evaluation protocols does not lead to better diarisation performance. Second, embedding extractors have not seen utterances in which multiple speakers exist. These inputs are inevitably present in speaker diarisation because of overlapped speech and speaker changes; they degrade the performance. To mitigate the first problem, we generate speaker verification evaluation protocols that better mimic the diarisation scenario. We propose two data augmentation techniques to alleviate the second problem, making embedding extractors aware of overlapped speech or speaker change input. One technique generates overlapped speech segments, and the other generates segments where two speakers utter sequentially. Extensive experimental results using three state-of-the-art speaker embedding extractors demonstrate that both proposed approaches are effective.

The goal of this work is to detect new spoken terms defined by users. While most previous works address Keyword Spotting (KWS) as a closed-set classification problem, this limits their transferability to unseen terms. The ability to define custom keywords has advantages in terms of user experience.In this paper, we propose a metric learning-based training strategy for user-defined keyword spotting. In particular, we make the following contributions: (1) we construct a large-scale keyword dataset with an existing speech corpus and propose a filtering method to remove data that degrade model training; (2) we propose a metric learning-based two-stage training strategy, and demonstrate that the proposed method improves the performance on the user-defined keyword spotting task by enriching their representations; (3) to facilitate the fair comparison in the user-defined KWS field, we propose unified evaluation protocol and metrics.Our proposed system does not require an incremental training on the user-defined keywords, and outperforms previous works by a significant margin on the Google Speech Commands dataset using the proposed as well as the existing metrics.

The objective of this work is to train noise-robust speaker embeddings adapted for speaker diarisation. Speaker embeddings play a crucial role in the performance of diarisation systems, but they often capture spurious information such as noise, adversely affecting performance. Our previous work has proposed an auto-encoder-based dimensionality reduction module to help remove the redundant information. However, they do not explicitly separate such information and have also been found to be sensitive to hyper-parameter values. To this end, we propose two contributions to overcome these issues: (i) a novel dimensionality reduction framework that can disentangle spurious information from the speaker embeddings; (ii) the use of speech activity vector to prevent the speaker code from representing the background noise. Through a range of experiments conducted on four datasets, our approach consistently demonstrates the state-of-the-art performance among models without system fusion.

The goal of this work is zero-shot text-to-speech synthesis, with speaking styles and voices learnt from facial characteristics. Inspired by the natural fact that people can imagine the voice of someone when they look at his or her face, we introduce a face-styled diffusion text-to-speech (TTS) model within a unified framework learnt from visible attributes, called Face-TTS. This is the first time that face images are used as a condition to train a TTS model.We jointly train cross-model biometrics and TTS models to preserve speaker identity between face images and generated speech segments. We also propose a speaker feature binding loss to enforce the similarity of the generated and the ground truth speech segments in speaker embedding space. Since the biometric information is extracted directly from the face image, our method does not require extra fine-tuning steps to generate speech from unseen and unheard speakers. We train and evaluate the model on the LRS3 dataset, an in-the-wild audio-visual corpus containing background noise and diverse speaking styles. The project page is https://facetts.github.io.

The goal of this work is to localize sound sources in visual scenes with a self-supervised approach. Contrastive learning in the context of sound source localization leverages the natural correspondence between audio and visual signals where the audio-visual pairs from the same source are assumed as positive, while randomly selected pairs are negatives. However, this approach brings in noisy correspondences; for example, positive audio and visual pair signals that may be unrelated to each other, or negative pairs that may contain semantically similar samples to the positive one. Our key contribution in this work is to show that using a less strict decision boundary in contrastive learning can alleviate the effect of noisy correspondences in sound source localization. We propose a simple yet effective approach by slightly modifying the contrastive loss with a negative margin. Extensive experimental results show that our approach gives on-par or better performance than the state-of-the-art methods. Furthermore, we demonstrate that the introduction of a negative margin to existing methods results in a consistent improvement in performance.

The objective of this work is to explore the learning of visually grounded speech models (VGS) from multilingual perspective. Bilingual VGS models are generally trained with an equal number of spoken captions from both languages. However, in reality, there can be an imbalance among the languages for the available spoken captions. Our key contribution in this work is to leverage the power of a high-resource language in a bilingual visually grounded speech model to improve the performance of a low-resource language. We introduce two methods to distill the knowledge of high-resource language into low-resource languages: (1) incorporating a strong pre-trained high-resource language encoder and (2) using semantically similar spoken captions. Our experiments show that combining these two approaches effectively enables the low-resource language to surpass the performances of monolingual and bilingual counterparts for cross-modal retrieval tasks.

Humans can easily perceive the direction of sound sources in a visual scene, termed sound source localization. Recent studies on learning-based sound source localization have mainly explored the problem from a localization perspective. However, prior arts and existing benchmarks do not account for a more important aspect of the problem, cross-modal semantic understanding, which is essential for genuine sound source localization. Cross-modal semantic understanding is important in understanding semantically mismatched audio-visual events, e.g., silent objects, or off-screen sounds. To account for this, we propose a cross-modal alignment task as a joint task with sound source localization to better learn the interaction between audio and visual modalities. Thereby, we achieve high localization performance with strong cross-modal semantic understanding. Our method outperforms the state-of-the-art approaches in both sound source localization and cross-modal retrieval. Our work suggests that jointly tackling both tasks is necessary to conquer genuine sound source localization.

The goal of this paper is to synthesise talking faces with controllable facial motions. To achieve this goal, we propose two key ideas. The first is to establish a canonical space where every face has the same motion patterns but different identities. The second is to navigate a multimodal motion space that only represents motion-related features while eliminating identity information. To disentangle identity and motion, we introduce an orthogonality constraint between the two different latent spaces. From this, our method can generate natural-looking talking faces with fully controllable facial attributes and accurate lip synchronisation. Extensive experiments demonstrate that our method achieves state-of-the-art results in terms of both visual quality and lip-sync score. To the best of our knowledge, we are the first to develop a talking face generation framework that can accurately manifest full target facial motions including lip, head pose, and eye movements in the generated video without any additional supervision beyond RGB video with audio.

Artefacts that differentiate spoofed from bona-fide utterances can reside in specific temporal or spectral intervals. Their reliable detection usually depends upon computationally demanding ensemble systems where each subsystem is tuned to some specific artefacts. We seek to develop an efficient, single system that can detect a broad range of different spoofing attacks without score-level ensembles. We propose a novel heterogeneous stacking graph attention layer that models artefacts spanning heterogeneous temporal and spectral intervals with a heterogeneous attention mechanism and a stack node. With a new max graph operation that involves a competitive mechanism and a new readout scheme, our approach, named AASIST, outperforms the current state-of-the-art by 20% relative. Even a lightweight variant, AASIST-L, with only 85k parameters, outperforms all competing systems.

Recognition of uncommon words such as names and technical terminology is important to understanding conversations in context. However, the ability to recognise such words remains a challenge in modern automatic speech recognition (ASR) systems.In this paper, we propose a simple but powerful ASR decoding method that can better recognise these uncommon keywords, which in turn enables better readability of the results. The method boosts the probabilities of given keywords in a beam search based on acoustic model predictions. The method does not require any training in advance.We demonstrate the effectiveness of our method on the LibriSpeeech test sets and also internal data of real-world conversations. Our method significantly boosts keyword accuracy on the test sets, while maintaining the accuracy of the other words, and as well as providing significant qualitative improvements. This method is applicable to other tasks such as machine translation, or wherever unseen and difficult keywords need to be recognised in beam search.

The objective of this work is effective speaker diarisation using multi-scale speaker embeddings. Typically, there is a trade-off between the ability to recognise short speaker segments and the discriminative power of the embedding, according to the segment length used for embedding extraction. To this end, recent works have proposed the use of multi-scale embeddings where segments with varying lengths are used. However, the scores are combined using a weighted summation scheme where the weights are fixed after the training phase, whereas the importance of segment lengths can differ within a single session.To address this issue, we present three key contributions in this paper: (1) we propose graph attention networks for multi-scale speaker diarisation; (2) we design scale indicators to utilise scale information of each embedding; (3) we adapt the attention-based aggregation to utilise a pre-computed affinity matrix from multi-scale embeddings.We demonstrate the effectiveness of our method in various datasets where the speaker confusion which constitutes the primary metric drops over 10% in average relative compared to the baseline.

In recent years, speaker recognition systems based on raw waveform inputs have received increasing attention.However, the performance of such systems are typically inferior to the state-of-the-art handcrafted feature-based counterparts, which demonstrate equal error rates under 1% on the popular VoxCeleb1 test set.This paper proposes a novel speaker recognition model based on raw waveform inputs.The model incorporates recent advances in machine learning and speaker verification, including the Res2Net backbone module and multi-layer feature aggregation.Our best model achieves an equal error rate of 0.89%, which is competitive with the state-of-the-art models based on handcrafted features, and outperforms the best model based on raw waveform inputs by a large margin.We also explore the application of the proposed model in the context of self-supervised learning framework.Our self-supervised model outperforms single phase-based existing works in this line of research.Finally, we show that self-supervised pre-training is effective for the semi-supervised scenario where we only have a small set of labelled training data, along with a larger set of unlabelled examples.

The goal of this work is to recognise phrases and sentences being spoken by a talking face, with or without the audio. Unlike previous works that have focussed on recognising a limited number of words or phrases, we tackle lip reading as an open-world problem - unconstrained natural language sentences, and in the wild videos. Our key contributions are: (1) we compare two models for lip reading, one using a CTC loss, and the other using a sequence-to-sequence loss. Both models are built on top of the transformer self-attention architecture; (2) we investigate to what extent lip reading is complementary to audio speech recognition, especially when the audio signal is noisy; (3) we introduce and publicly release a new dataset for audio-visual speech recognition, LRS2-BBC, consisting of thousands of natural sentences from British television. The models that we train surpass the performance of all previous work on a lip reading benchmark dataset by a significant margin.

The goal of this work is to investigate the performance of popular speaker recognition models on speech segments from movies, where often actors intentionally disguise their voice to play a character. We make the following three contributions: (i) We collect a novel, challenging speaker recognition dataset called VoxMovies, with speech for 856 identities from almost 4000 movie clips. VoxMovies contains utterances with varying emotion, accents and background noise, and therefore comprises an entirely different domain to the interview-style, emotionally calm utterances in current speaker recognition datasets such as VoxCeleb; (ii) We provide a number of domain adaptation evaluation sets, and benchmark the performance of state-of-the-art speaker recognition models on these evaluation pairs. We demonstrate that both speaker verification and identification performance drops steeply on this new data, showing the challenge in transferring models across domains; and finally (iii) We show that simple domain adaptation paradigms improve performance, but there is still large room for improvement.

This work presents a novel back-end framework for speaker verification using graph attention networks. Segment-wise speaker embeddings extracted from multiple crops within an utterance are interpreted as node representations of a graph. The proposed framework inputs segment-wise speaker embeddings from an enrollment and a test utterance and directly outputs a similarity score. We first construct a graph using segment-wise speaker embeddings and then input these to graph attention networks. After a few graph attention layers with residual connections, each node is projected into a one-dimensional space using affine transform, followed by a readout operation resulting in a scalar similarity score. To enable successful adaptation for speaker verification, we propose techniques such as separating trainable weights for attention map calculations between segment-wise speaker embeddings from different utterances. The effectiveness of the proposed framework is validated using three different speaker embedding extractors trained with different architectures and objective functions. Experimental results demonstrate consistent improvement over various baseline back-end classifiers, with an average equal error rate improvement of 20% over the cosine similarity back-end without test time augmentation.

The VoxCeleb Speaker Recognition Challenge (VoxSRC) at Interspeech 2020 offers a challenging evaluation for speaker recognition systems, which includes celebrities playing different parts in movies. The goal of this work is robust speaker recognition of utterances recorded in these challenging environments. We utilise variants of the popular ResNet architecture for speaker recognition and perform extensive experiments using a range of loss functions and training parameters. To this end, we optimise an efficient training framework that allows powerful models to be trained with limited time and resources. Our trained models demonstrate improvements over most existing works with lighter models and a simple pipeline. The paper shares the lessons learned from our participation in the challenge.

In this work, we propose an overlapped speech detection system trained as a three-class classifier. Unlike conventional systems that perform binary classification as to whether or not a frame contains overlapped speech, the proposed approach classifies into three classes: non-speech, single speaker speech, and overlapped speech. By training a network with the more detailed label definition, the model can learn a better notion on deciding the number of speakers included in a given frame. A convolutional recurrent neural network architecture is explored to benefit from both convolutional layer's capability to model local patterns and recurrent layer's ability to model sequential information. The proposed overlapped speech detection model establishes a state-of-the-art performance with a precision of 0.6648 and a recall of 0.3222 on the DIHARD II evaluation set, showing a 20% increase in recall along with higher precision. In addition, we also introduce a simple approach to utilize the proposed overlapped speech detection model for speaker diarization which ranked third place in the Track 1 of the DIHARD III challenge.

In this work, we present a novel audio-visual dataset for active speaker detection in the wild.A speaker is considered active when his or her face is visible and the voice is audible simultaneously.Although active speaker detection is a crucial pre-processing step for many audio-visual tasks, there is no existing dataset of natural human speech to evaluate the performance of active speaker detection.We therefore curate the Active Speakers in the Wild (ASW) dataset which contains videos and co-occurring speech segments with dense speech activity labels.Videos and timestamps of audible segments are parsed and adopted from VoxConverse, an existing speaker diarisation dataset that consists of videos in the wild.Face tracks are extracted from the videos and active segments are annotated based on the timestamps of VoxConverse in a semi-automatic way.Two reference systems, a self-supervised system and a fully supervised one, are evaluated on the dataset to provide the baseline performances of ASW.Cross-domain evaluation is conducted in order to show the negative effect of dubbed videos in the training data.

The goal of this paper is to adapt speaker embeddings for solving the problem of speaker diarisation. The quality of speaker embeddings is paramount to the performance of speaker diarisation systems. Despite this, prior works in the field have directly used embeddings designed only to be effective on the speaker verification task. In this paper, we propose three techniques that can be used to better adapt the speaker embeddings for diarisation: dimensionality reduction, attention-based embedding aggregation, and non-speech clustering. A wide range of experiments is performed on various challenging datasets. The results demonstrate that all three techniques contribute positively to the performance of the diarisation system achieving an average relative improvement of 25.07% in terms of diarisation error rate over the baseline.

The goal of this work is to train effective representations for keyword spotting via metric learning. Most existing works address keyword spotting as a closed-set classification problem, where both target and non-target keywords are predefined. Therefore, prevailing classifier-based keyword spot-ting systems perform poorly on non-target sounds which are unseen during the training stage, causing high false alarm rates in real-world scenarios. In reality, keyword spotting is a detection problem where predefined target keywords are detected from a variety of unknown sounds. This shares many similarities to metric learning problems in that the unseen and unknown non-target sounds must be clearly differentiated from the target keywords. However, a key difference is that the target keywords are known and predefined. To this end, we propose a new method based on metric learning that maximises the distance between target and non-target key-words, but also learns per-class weights for target keywords as in classification objectives. Experiments on the Google Speech Commands dataset show that our method significantly reduces false alarms to unseen non-target keywords, while maintaining the overall classification accuracy.

The objective of this work is speaker diarisation of speech recordings `in the wild'. The ability to determine speech segments is a crucial part of diarisation systems, accounting for a large proportion of errors. In this paper, we present a simple but effective solution for speech activity detection based on the speaker embeddings. In particular, we discover that the norm of the speaker embedding is an extremely effective indicator of speech activity. The method does not require an independent model for speech activity detection, therefore allows speaker diarisation to be performed using a unified representation for both speaker modelling and speech activity detection. We perform a number of experiments on in-house and public datasets, in which our method outperforms popular baselines.

The recently proposed self-attentive pooling (SAP) has shown good performance in several speaker recognition systems. In SAP systems, the context vector is trained end-to-end together with the feature extractor, where the role of context vector is to select the most discriminative frames for speaker recognition. However, the SAP underperforms compared to the temporal average pooling (TAP) baseline in some settings, which implies that the attention is not learnt effectively in end-to-end training. To tackle this problem, we introduce strategies for training the attention mechanism in a supervised manner, which learns the context vector using classified samples. With our proposed methods, context vector can be boosted to select the most informative frames. We show that our method outperforms existing methods in various experimental settings including short utterance speaker recognition, and achieves competitive performance over the existing baselines on the VoxCeleb datasets.

The goal of this paper is text-independent speaker verification where utterances come from `in the wild' videos and may contain irrelevant signal. While speaker verification is naturally a pair-wise problem, existing methods to produce the speaker embeddings are instance-wise. In this paper, we propose Cross Attentive Pooling (CAP) that utilises the context information across the reference-query pair to generate utterance-level embeddings that contain the most discriminative information for the pair-wise matching problem. Experiments are performed on the VoxCeleb dataset in which our method outperforms comparable pooling strategies.

The goal of this work is to train strong models for visual speech recognition without requiring human annotated ground truth data. We achieve this by distilling from an Automatic Speech Recognition (ASR) model that has been trained on a large-scale audio-only corpus. We use a cross-modal distillation method that combines Connectionist Temporal Classification (CTC) with a frame-wise cross-entropy loss. Our contributions are fourfold: (i) we show that ground truth transcriptions are not necessary to train a lip reading system; (ii) we show how arbitrary amounts of unlabelled video data can be leveraged to improve performance; (iii) we demonstrate that distillation significantly speeds up training; and, (iv) we obtain state-of-the-art results on the challenging LRS2 and LRS3 datasets for training only on publicly available data.

Content and style representations have been widely studied in the field of style transfer. In this paper, we propose a new loss function using speaker content representation for audio source separation, and we call it speaker representation loss. The objective is to extract the target speaker voice from the noisy input and also remove it from the residual components. Compared to the conventional spectral reconstruction, our proposed framework maximizes the use of target speaker information by minimizing the distance between the speaker representations of reference and source separation output. We also propose triplet speaker representation loss as an additional criterion to remove the target speaker information from residual spectrogram output. VoiceFilter framework is adopted to evaluate source separation performance using the VCTK database, and we achieved improved performances compared to the baseline loss function without any additional network parameters.

The objective of this paper is to learn representations of speaker identity without access to manually annotated data. To do so, we develop a self-supervised learning objective that exploits the natural cross-modal synchrony between faces and audio in video. The key idea behind our approach is to tease apart-without annotation-the representations of linguistic content and speaker identity. We construct a two-stream architecture which: (1) shares low-level features common to both representations; and (2) provides a natural mechanism for explicitly disentangling these factors, offering the potential for greater generalisation to novel combinations of content and identity and ultimately producing speaker identity representations that are more robust. We train our method on a large-scale audio-visual dataset of talking heads `in the wild', and demonstrate its efficacy by evaluating the learned speaker representations for standard speaker recognition performance.

The goal of this work is to train models that can identify a spoken language just by interpreting the speaker's lip movements.Our contributions are the following: (i) we show that models can learn to discriminate among 14 different languages using only visual speech information; (ii) we compare different designs in sequence modelling and utterance-level aggregation in order to determine the best architecture for this task; (iii) we investigate the factors that contribute discriminative cues and show that our model indeed solves the problem by finding temporal patterns in mouth movements and not by exploiting spurious correlations.We demonstrate this further by evaluating our models on challenging examples from bilingual speakers.

The objective of this paper is to separate a target speaker's speech from a mixture of two speakers using a deep audio-visual speech separation network. Unlike previous works that used lip movement on video clips or pre-enrolled speaker information as an auxiliary conditional feature, we use a single face image of the target speaker. In this task, the conditional feature is obtained from facial appearance in cross-modal biometric task, where audio and visual identity representations are shared in latent space. Learnt identities from facial images enforce the network to isolate matched speakers and extract the voices from mixed speech. It solves the permutation problem caused by swapped channel outputs, frequently occurred in speech separation tasks. The proposed method is far more practical than video-based speech separation since user profile images are readily available on many platforms. Also, unlike speaker-aware separation methods, it is applicable on separation with unseen speakers who have never been enrolled before. We show strong qualitative and quantitative results on challenging real-world examples.

The objective of this paper is 'open-set' speaker recognition of unseen speakers, where ideal embeddings should be able to condense information into a compact utterance-level representation that has small intra-speaker and large inter-speaker distance. A popular belief in speaker recognition is that networks trained with classification objectives outperform metric learning methods. In this paper, we present an extensive evaluation of most popular loss functions for speaker recognition on the VoxCeleb dataset. We demonstrate that the vanilla triplet loss shows competitive performance compared to classification-based losses, and those trained with our proposed metric learning objective outperform state-of-the-art methods.

The goal of this paper is speaker diarisation of videos collected 'in the wild'. We make three key contributions. First, we propose an automatic audio-visual diarisation method for YouTube videos. Our method consists of active speaker detection using audio-visual methods and speaker verification using self-enrolled speaker models. Second, we integrate our method into a semi-automatic dataset creation pipeline which significantly reduces the number of hours required to annotate videos with diarisation labels. Finally, we use this pipeline to create a large-scale diarisation dataset called VoxConverse, collected from 'in the wild' videos, which we will release publicly to the research community. Our dataset consists of overlapping speech, a large and diverse speaker pool, and challenging background conditions.

The goal of this work is to train discriminative cross-modal embeddings without access to manually annotated data. Recent advances in self-supervised learning have shown that effective representations can be learnt from natural cross-modal synchrony. We build on earlier work to train embeddings that are more discriminative for uni-modal downstream tasks. To this end, we propose a novel training strategy that not only optimises metrics across modalities, but also enforces intra-class feature separation within each of the modalities. The effectiveness of the method is demonstrated on two downstream tasks: lip reading using the features trained on audio-visual synchronisation, and speaker recognition using the features trained for cross-modal biometric matching. The proposed method outperforms state-of-the-art self-supervised baselines by a signficant margin.

The objective of this work is speaker recognition under noisy and unconstrained conditions. We make two key contributions. First, we introduce a very large-scale audio-visual dataset collected from open source media using a fully automated pipeline. Most existing datasets for speaker identification contain samples obtained under quite constrained conditions, and usually require manual annotations, hence are limited in size. We propose a pipeline based on computer vision techniques to create the dataset from open-source media. Our pipeline involves obtaining videos from YouTube; performing active speaker verification using a two-stream synchronization Convolutional Neural Network (CNN), and confirming the identity of the speaker using CNN based facial recognition. We use this pipeline to curate VoxCeleb which contains contains over a million ‘real-world’ utterances from over 6000 speakers. This is several times larger than any publicly available speaker recognition dataset. Second, we develop and compare different CNN architectures with various aggregation methods and training loss functions that can effectively recognise identities from voice under various conditions. The models trained on our dataset surpass the performance of previous works by a significant margin.

This paper proposes a new strategy for learning powerful cross-modal embeddings for audio-to-video synchronisation. Here, we set up the problem as one of cross-modal retrieval, where the objective is to find the most relevant audio segment given a short video clip. The method builds on the recent advances in learning representations from cross-modal self-supervision. The main contributions of this paper are as follows: (1) we propose a new learning strategy where the embeddings are learnt via a multi-way matching problem, as opposed to a binary classification (matching or non-matching) problem as proposed by recent papers; (2) we demonstrate that performance of this method far exceeds the existing baselines on the synchronisation task; (3) we use the learnt embeddings for visual speech recognition in self-supervision, and show that the performance matches the representations learnt end-to-end in a fully-supervised manner.

The objective of this paper is speaker recognition `in the wild' - where utterances may be of variable length and also contain irrelevant signals. Crucial elements in the design of deep networks for this task are the type of trunk (frame level) network, and the method of temporal aggregation. We propose a powerful speaker recognition deep network, using a `thin-ResNet' trunk architecture, and a dictionary-based NetVLAD or GhostVLAD layer to aggregate features across time, that can be trained end-to-end. We show that our network achieves state of the art performance by a significant margin on the VoxCeleb1 test set for speaker recognition, whilst requiring fewer parameters than previous methods. We also investigate the effect of utterance length on performance, and conclude that for `in the wild' data, a longer length is beneficial.

Our objective is an audio-visual model for separating a single speaker from a mixture of sounds such as other speakers and background noise. Moreover, we wish to hear the speaker even when the visual cues are temporarily absent due to occlusion. To this end we introduce a deep audio-visual speech enhancement network that is able to separate a speaker's voice by conditioning on both the speaker's lip movements and/or a representation of their voice. The voice representation can be obtained by either (i) enrollment, or (ii) by self-enrollment -- learning the representation on-the-fly given sufficient unobstructed visual input. The model is trained by blending audios, and by introducing artificial occlusions around the mouth region that prevent the visual modality from dominating. The method is speaker-independent, and we demonstrate it on real examples of speakers unheard (and unseen) during training. The method also improves over previous models in particular for cases of occlusion in the visual modality.

The goal of this work is to determine who spoke when' in real-world meetings.The method takes surround-view video and single or multi-channel audio as inputs, and generates robust diarisation outputs.To achieve this, we propose a novel iterative approach that first enrolls speaker models using audio-visual correspondence, then uses the enrolled models together with the visual information to determine the active speaker.We show strong quantitative and qualitative performance on a dataset of real-world meetings.The method is also evaluated on the public AMI meeting corpus, on which we demonstrate results that exceed all comparable methods.We also show that beamforming can be used together with the video to further improve the performance when multi-channel audio is available.

We describe a method for generating a video of a talking face. The method takes still images of the target face and an audio speech segment as inputs, and generates a video of the target face lip synched with the audio. The method runs in real time and is applicable to faces and audio not seen at training time. To achieve this we develop an encoder–decoder convolutional neural network (CNN) model that uses a joint embedding of the face and audio to generate synthesised talking face video frames. The model is trained on unlabelled videos using cross-modal self-supervision. We also propose methods to re-dub videos by visually blending the generated face into the source video frame using a multi-stream CNN model.

Our goal is to isolate individual speakers from multi-talker simultaneous speech in videos. Existing works in this area have focussed on trying to separate utterances from known speakers in controlled environments. In this paper, we propose a deep audio-visual speech enhancement network that is able to separate a speaker's voice given lip regions in the corresponding video, by predicting both the magnitude and the phase of the target signal. The method is applicable to speakers unheard and unseen during training, and for unconstrained environments. We demonstrate strong quantitative and qualitative results, isolating extremely challenging real-world examples.

The goal of this paper is to develop state-of-the-art models for lip reading -- visual speech recognition. We develop three architectures and compare their accuracy and training times: (i) a recurrent model using LSTMs; (ii) a fully convolutional model; and (iii) the recently proposed transformer model. The recurrent and fully convolutional models are trained with a Connectionist Temporal Classification loss and use an explicit language model for decoding, the transformer is a sequence-to-sequence model. Our best performing model improves the state-of-the-art word error rate on the challenging BBC-Oxford Lip Reading Sentences 2 (LRS2) benchmark dataset by over 20 percent. As a further contribution we investigate the fully convolutional model when used for online (real time) lip reading of continuous speech, and show that it achieves high performance with low latency.

The objective of this paper is speaker recognition under noisy and unconstrained conditions. We make two key contributions. First, we introduce a very large-scale audio-visual speaker recognition dataset collected from open-source media. Using a fully automated pipeline, we curate VoxCeleb2 which contains over a million utterances from over 6,000 speakers. This is several times larger than any publicly available speaker recognition dataset. Second, we develop and compare Convolutional Neural Network (CNN) models and training strategies that can effectively recognise identities from voice under various conditions. The models trained on the VoxCeleb2 dataset surpass the performance of previous works on a benchmark dataset by a significant margin.

The goal of this work is to recognise phrases and sentences being spoken by a talking face, with or without the audio. Unlike previous works that have focussed on recognising a limited number of words or phrases, we tackle lip reading as an open-world problem - unconstrained natural language sentences, and in the wild videos. Our key contributions are: (1) a `Watch, Listen, Attend and Spell' (WLAS) network that learns to transcribe videos of mouth motion to characters; (2) a curriculum learning strategy to accelerate training and to reduce overfitting; (3) a `Lip Reading Sentences' (LRS) dataset for visual speech recognition, consisting of over 100,000 natural sentences from British television. The WLAS model trained on the LRS dataset surpasses the performance of all previous work on standard lip reading benchmark datasets, often by a significant margin. This lip reading performance beats a professional lip reader on videos from BBC television, and we also demonstrate that if audio is available, then visual information helps to improve speech recognition performance.

Most existing datasets for speaker identification contain samples obtained under quite constrained conditions, and are usually hand-annotated, hence limited in size. The goal of this paper is to generate a large scale text-independent speaker identification dataset collected 'in the wild'. We make two contributions. First, we propose a fully automated pipeline based on computer vision techniques to create the dataset from open-source media. Our pipeline involves obtaining videos from YouTube; performing active speaker verification using a two-stream synchronization Convolutional Neural Network (CNN), and confirming the identity of the speaker using CNN based facial recognition. We use this pipeline to curate VoxCeleb which contains hundreds of thousands of 'real world' utterances for over 1,000 celebrities. Our second contribution is to apply and compare various state of the art speaker identification techniques on our dataset to establish baseline performance. We show that a CNN based architecture obtains the best performance for both identification and verification.