Philip J. B. Jackson

Faegheh Sardari, Armin Mustafa, Philip J. B. Jackson et al.

We present PAT, a transformer-based network that learns complex temporal co-occurrence action dependencies in a video by exploiting multi-scale temporal features. In existing methods, the self-attention mechanism in transformers loses the temporal positional information, which is essential for robust action detection. To address this issue, we (i) embed relative positional encoding in the self-attention mechanism and (ii) exploit multi-scale temporal relationships by designing a novel non hierarchical network, in contrast to the recent transformer-based approaches that use a hierarchical structure. We argue that joining the self-attention mechanism with multiple sub-sampling processes in the hierarchical approaches results in increased loss of positional information. We evaluate the performance of our proposed approach on two challenging dense multi-label benchmark datasets, and show that PAT improves the current state-of-the-art result by 1.1% and 0.6% mAP on the Charades and MultiTHUMOS datasets, respectively, thereby achieving the new state-of-the-art mAP at 26.5% and 44.6%, respectively. We also perform extensive ablation studies to examine the impact of the different components of our proposed network.

Davide Berghi, Philip J. B. Jackson

Sound event localization and detection (SELD) involves predicting active sound event classes over time while estimating their positions. The localization subtask in SELD is usually treated as a direction of arrival estimation problem, ignoring source distance. Only recently, SELD was extended to 3D by incorporating distance estimation, enabling the prediction of sound event positions in 3D space (3D SELD). However, existing methods lack input features specifically designed for distance estimation. We address this gap by introducing two novel reverberation-based feature formats: one using the direct-to-reverberant ratio (DRR) and another leveraging signal autocorrelation to capture early reflections. We extensively evaluate and benchmark these features on the STARSS23 dataset, combining them with established SELD features for sound event detection (SED) and direction-of-arrival estimation (DOAE), and testing across different network architectures. Our proposed features, applicable to both FOA and MIC formats, achieve state-of-the-art distance estimation, enhancing overall 3D SELD performance.

Faegheh Sardari, Armin Mustafa, Philip J. B. Jackson et al.

Weakly supervised audio-visual video parsing (AVVP) methods aim to detect audible-only, visible-only, and audible-visible events using only video-level labels. Existing approaches tackle this by leveraging unimodal and cross-modal contexts. However, we argue that while cross-modal learning is beneficial for detecting audible-visible events, in the weakly supervised scenario, it negatively impacts unaligned audible or visible events by introducing irrelevant modality information. In this paper, we propose CoLeaF, a novel learning framework that optimizes the integration of cross-modal context in the embedding space such that the network explicitly learns to combine cross-modal information for audible-visible events while filtering them out for unaligned events. Additionally, as videos often involve complex class relationships, modelling them improves performance. However, this introduces extra computational costs into the network. Our framework is designed to leverage cross-class relationships during training without incurring additional computations at inference. Furthermore, we propose new metrics to better evaluate a method's capabilities in performing AVVP. Our extensive experiments demonstrate that CoLeaF significantly improves the state-of-the-art results by an average of 1.9% and 2.4% F-score on the LLP and UnAV-100 datasets, respectively.

Davide Berghi, Philip J. B. Jackson

This report describes our systems submitted for the DCASE2024 Task 3 challenge: Audio and Audiovisual Sound Event Localization and Detection with Source Distance Estimation (Track B). Our main model is based on the audio-visual (AV) Conformer, which processes video and audio embeddings extracted with ResNet50 and with an audio encoder pre-trained on SELD, respectively. This model outperformed the audio-visual baseline of the development set of the STARSS23 dataset by a wide margin, halving its DOAE and improving the F1 by more than 3x. Our second system performs a temporal ensemble from the outputs of the AV-Conformer. We then extended the model with features for distance estimation, such as direct and reverberant signal components extracted from the omnidirectional audio channel, and depth maps extracted from the video frames. While the new system improved the RDE of our previous model by about 3 percentage points, it achieved a lower F1 score. This may be caused by sound classes that rarely appear in the training set and that the more complex system does not detect, as analysis can determine. To overcome this problem, our fourth and final system consists of an ensemble strategy combining the predictions of the other three. Many opportunities to refine the system and training strategy can be tested in future ablation experiments, and likely achieve incremental performance gains for this audio-visual task.

Davide Berghi, Philip J. B. Jackson

Conventional audio-visual approaches for active speaker detection (ASD) typically rely on visually pre-extracted face tracks and the corresponding single-channel audio to find the speaker in a video. Therefore, they tend to fail every time the face of the speaker is not visible. We demonstrate that a simple audio convolutional recurrent neural network (CRNN) trained with spatial input features extracted from multichannel audio can perform simultaneous horizontal active speaker detection and localization (ASDL), independently of the visual modality. To address the time and cost of generating ground truth labels to train such a system, we propose a new self-supervised training pipeline that embraces a ``student-teacher'' learning approach. A conventional pre-trained active speaker detector is adopted as a ``teacher'' network to provide the position of the speakers as pseudo-labels. The multichannel audio ``student'' network is trained to generate the same results. At inference, the student network can generalize and locate also the occluded speakers that the teacher network is not able to detect visually, yielding considerable improvements in recall rate. Experiments on the TragicTalkers dataset show that an audio network trained with the proposed self-supervised learning approach can exceed the performance of the typical audio-visual methods and produce results competitive with the costly conventional supervised training. We demonstrate that improvements can be achieved when minimal manual supervision is introduced in the learning pipeline. Further gains may be sought with larger training sets and integrating vision with the multichannel audio system.

Davide Berghi, Philip J. B. Jackson

In this study, we address the multimodal task of stereo sound event localization and detection with source distance estimation (3D SELD) in regular video content. 3D SELD is a complex task that combines temporal event classification with spatial localization, requiring reasoning across spatial, temporal, and semantic dimensions. The last is arguably the most challenging to model. Traditional SELD approaches typically rely on multichannel input, limiting their capacity to benefit from large-scale pre-training due to data constraints. To overcome this, we enhance a standard SELD architecture with semantic information by integrating pre-trained, contrastive language-aligned models: CLAP for audio and OWL-ViT for visual inputs. These embeddings are incorporated into a modified Conformer module tailored for multimodal fusion, which we refer to as the Cross-Modal Conformer. We perform an ablation study on the development set of the DCASE2025 Task3 Stereo SELD Dataset to assess the individual contributions of the language-aligned models and benchmark against the DCASE Task 3 baseline systems. Additionally, we detail the curation process of large synthetic audio and audio-visual datasets used for model pre-training. These datasets were further expanded through left-right channel swapping augmentation. Our approach, combining extensive pre-training, model ensembling, and visual post-processing, achieved second rank in the DCASE 2025 Challenge Task 3 (Track B), underscoring the effectiveness of our method. Future work will explore the modality-specific contributions and architectural refinements.

Davide Berghi, Philip J. B. Jackson

This report presents our systems submitted to the audio-only and audio-visual tracks of the DCASE2025 Task 3 Challenge: Stereo Sound Event Localization and Detection (SELD) in Regular Video Content. SELD is a complex task that combines temporal event classification with spatial localization, requiring reasoning across spatial, temporal, and semantic dimensions. The last is arguably the most challenging to model. Traditional SELD architectures rely on multichannel input, which limits their ability to leverage large-scale pre-training due to data constraints. To address this, we enhance standard SELD architectures with semantic information by integrating pre-trained, contrastive language-aligned models: CLAP for audio and OWL-ViT for visual inputs. These embeddings are incorporated into a modified Conformer module tailored for multimodal fusion, which we refer to as the Cross-Modal Conformer. Additionally, we incorporate autocorrelation-based acoustic features to improve distance estimation. We pre-train our models on curated synthetic audio and audio-visual datasets and apply a left-right channel swapping augmentation to further increase the training data. Both our audio-only and audio-visual systems substantially outperform the challenge baselines on the development set, demonstrating the effectiveness of our strategy. Performance is further improved through model ensembling and a visual post-processing step based on human keypoints. Future work will investigate the contribution of each modality and explore architectural variants to further enhance results.

Tony Alex, Wish Suharitdamrong, Sara Atito et al.

Integration of audio perception into large language models (LLMs) is an emerging research area for enabling machine listening applications, yet efficient transfer of rich audio semantics from audio encoders to LLMs remains underexplored. The most widely used integration paradigm projects the audio encoder output tokens into the LLM input space (e.g., via an MLP or a Q-Former), then prepends or inserts them to the text tokens. We refer to this generic scheme as Prepend to the LLM's input token space (PLITS) integration. We propose an efficient alternative, Lightweight Audio LLM Integration (LAL). LAL introduces audio representations solely via the attention mechanism within different layers of the LLM, bypassing its feedforward module. LAL encodes rich audio semantics at an appropriate level of abstraction for integration into different blocks of LLMs. Our design significantly reduces computational overhead compared to existing integration approaches. Observing with Whisper that the speech encoder benefits from PLITS integration, we propose an audio encoder aware approach for efficiently Probing Audio encoders via LLM (PAL), which employs PLITS integration for Whisper and LAL for general audio encoders. Under an identical training curriculum, LAL consistently maintains performance or outperforms existing integration approaches across multiple base LLMs and tasks. For general audio tasks, LAL improvement is up to 30% over a strong PLITS baseline while reducing memory usage by up to 64.1% and increasing throughput by up to 247.5%. Furthermore, for general audio-music-speech LLM, PAL performs on par with a fully PLITS integration-based system but with substantially improved computational and memory efficiency. Project page: https://ta012.github.io/PAL/

Faegheh Sardari, Armin Mustafa, Philip J. B. Jackson et al.

Dense action detection involves detecting multiple co-occurring actions while action classes are often ambiguous and represent overlapping concepts. We argue that handling the dual challenge of temporal and class overlaps is too complex to effectively be tackled by a single network. To address this, we propose to decompose the task of detecting dense ambiguous actions into detecting dense, unambiguous sub-concepts that form the action classes (i.e., action entities and action motions), and assigning these sub-tasks to distinct sub-networks. By isolating these unambiguous concepts, the sub-networks can focus exclusively on resolving a single challenge, dense temporal overlaps. Furthermore, simultaneous actions in a video often exhibit interrelationships, and exploiting these relationships can improve the method performance. However, current dense action detection networks fail to effectively learn these relationships due to their reliance on binary cross-entropy optimization, which treats each class independently. To address this limitation, we propose providing explicit supervision on co-occurring concepts during network optimization through a novel language-guided contrastive learning loss. Our extensive experiments demonstrate the superiority of our approach over state-of-the-art methods, achieving substantial improvements of 3.8% and 1.7% on average across all metrics on the challenging benchmark datasets, Charades and MultiTHUMOS.

Faegheh Sardari, Armin Mustafa, Philip J. B. Jackson et al.

Unlike the sparse label action detection task, where a single action occurs in each timestamp of a video, in a dense multi-label scenario, actions can overlap. To address this challenging task, it is necessary to simultaneously learn (i) temporal dependencies and (ii) co-occurrence action relationships. Recent approaches model temporal information by extracting multi-scale features through hierarchical transformer-based networks. However, the self-attention mechanism in transformers inherently loses temporal positional information. We argue that combining this with multiple sub-sampling processes in hierarchical designs can lead to further loss of positional information. Preserving this information is essential for accurate action detection. In this paper, we address this issue by proposing a novel transformer-based network that (a) employs a non-hierarchical structure when modelling different ranges of temporal dependencies and (b) embeds relative positional encoding in its transformer layers. Furthermore, to model co-occurrence action relationships, current methods explicitly embed class relations into the transformer network. However, these approaches are not computationally efficient, as the network needs to compute all possible pair action class relations. We also overcome this challenge by introducing a novel learning paradigm that allows the network to benefit from explicitly modelling temporal co-occurrence action dependencies without imposing their additional computational costs during inference. We evaluate the performance of our proposed approach on two challenging dense multi-label benchmark datasets and show that our method improves the current state-of-the-art results.

Davide Berghi, Philip J. B. Jackson

Object-based audio production requires the positional metadata to be defined for each point-source object, including the key elements in the foreground of the sound scene. In many media production use cases, both cameras and microphones are employed to make recordings, and the human voice is often a key element. In this research, we detect and locate the active speaker in the video, facilitating the automatic extraction of the positional metadata of the talker relative to the camera's reference frame. With the integration of the visual modality, this study expands upon our previous investigation focused solely on audio-based active speaker detection and localization. Our experiments compare conventional audio-visual approaches for active speaker detection that leverage monaural audio, our previous audio-only method that leverages multichannel recordings from a microphone array, and a novel audio-visual approach integrating vision and multichannel audio. We found the role of the two modalities to complement each other. Multichannel audio, overcoming the problem of visual occlusions, provides a double-digit reduction in detection error compared to audio-visual methods with single-channel audio. The combination of multichannel audio and vision further enhances spatial accuracy, leading to a four-percentage point increase in F1 score on the Tragic Talkers dataset. Future investigations will assess the robustness of the model in noisy and highly reverberant environments, as well as tackle the problem of off-screen speakers.

Hanne Stenzel, Davide Berghi, Marco Volino et al.

As audio-visual systems increasingly bring immersive and interactive capabilities into our work and leisure activities, so the need for naturalistic test material grows. New volumetric datasets have captured high-quality 3D video, but accompanying audio is often neglected, making it hard to test an integrated bimodal experience. Designed to cover diverse sound types and features, the presented volumetric dataset was constructed from audio and video studio recordings of scenes to yield forty short action sequences. Potential uses in technical and scientific tests are discussed.

Davide Berghi, Hanne Stenzel, Marco Volino et al.

Immersive audio-visual perception relies on the spatial integration of both auditory and visual information which are heterogeneous sensing modalities with different fields of reception and spatial resolution. This study investigates the perceived coherence of audiovisual object events presented either centrally or peripherally with horizontally aligned/misaligned sound. Various object events were selected to represent three acoustic feature classes. Subjective test results in a simulated virtual environment from 18 participants indicate a wider capture region in the periphery, with an outward bias favoring more lateral sounds. Centered stimulus results support previous findings for simpler scenes.

Luca Remaggi, Philip J. B. Jackson, Wenwu Wang

Typical methods for binaural source separation consider only the direct sound as the target signal in a mixture. However, in most scenarios, this assumption limits the source separation performance. It is well known that the early reflections interact with the direct sound, producing acoustic effects at the listening position, e.g. the so-called comb filter effect. In this article, we propose a novel source separation model, that utilizes both the direct sound and the first early reflection information to model the comb filter effect. This is done by observing the interaural phase difference obtained from the time-frequency representation of binaural mixtures. Furthermore, a method is proposed to model the interaural coherence of the signals. Including information related to the sound multipath propagation, the performance of the proposed separation method is improved with respect to the baselines that did not use such information, as illustrated by using binaural recordings made in four rooms, having different sizes and reverberation times.

Qiuqiang Kong, Yong Xu, Wenwu Wang et al.

Single-channel signal separation and deconvolution aims to separate and deconvolve individual sources from a single-channel mixture and is a challenging problem in which no prior knowledge of the mixing filters is available. Both individual sources and mixing filters need to be estimated. In addition, a mixture may contain non-stationary noise which is unseen in the training set. We propose a synthesizing-decomposition (S-D) approach to solve the single-channel separation and deconvolution problem. In synthesizing, a generative model for sources is built using a generative adversarial network (GAN). In decomposition, both mixing filters and sources are optimized to minimize the reconstruction error of the mixture. The proposed S-D approach achieves a peak-to-noise-ratio (PSNR) of 18.9 dB and 15.4 dB in image inpainting and completion, outperforming a baseline convolutional neural network PSNR of 15.3 dB and 12.2 dB, respectively and achieves a PSNR of 13.2 dB in source separation together with deconvolution, outperforming a convolutive non-negative matrix factorization (NMF) baseline of 10.1 dB.

Luca Remaggi, Philip J. B. Jackson, Philip Coleman et al.

Acoustic reflector localization is an important issue in audio signal processing, with direct applications in spatial audio, scene reconstruction, and source separation. Several methods have recently been proposed to estimate the 3D positions of acoustic reflectors given room impulse responses (RIRs). In this article, we categorize these methods as "image-source reversion", which localizes the image source before finding the reflector position, and "direct localization", which localizes the reflector without intermediate steps. We present five new contributions. First, an onset detector, called the clustered dynamic programming projected phase-slope algorithm, is proposed to automatically extract the time of arrival for early reflections within the RIRs of a compact microphone array. Second, we propose an image-source reversion method that uses the RIRs from a single loudspeaker. It is constructed by combining an image source locator (the image source direction and range (ISDAR) algorithm), and a reflector locator (using the loudspeaker-image bisection (LIB) algorithm). Third, two variants of it, exploiting multiple loudspeakers, are proposed. Fourth, we present a direct localization method, the ellipsoid tangent sample consensus (ETSAC), exploiting ellipsoid properties to localize the reflector. Finally, systematic experiments on simulated and measured RIRs are presented, comparing the proposed methods with the state-of-the-art. ETSAC generates errors lower than the alternative methods compared through our datasets. Nevertheless, the ISDAR-LIB combination performs well and has a run time 200 times faster than ETSAC.

Yong Xu, Qiang Huang, Wenwu Wang et al.

Environmental audio tagging aims to predict only the presence or absence of certain acoustic events in the interested acoustic scene. In this paper we make contributions to audio tagging in two parts, respectively, acoustic modeling and feature learning. We propose to use a shrinking deep neural network (DNN) framework incorporating unsupervised feature learning to handle the multi-label classification task. For the acoustic modeling, a large set of contextual frames of the chunk are fed into the DNN to perform a multi-label classification for the expected tags, considering that only chunk (or utterance) level rather than frame-level labels are available. Dropout and background noise aware training are also adopted to improve the generalization capability of the DNNs. For the unsupervised feature learning, we propose to use a symmetric or asymmetric deep de-noising auto-encoder (sDAE or aDAE) to generate new data-driven features from the Mel-Filter Banks (MFBs) features. The new features, which are smoothed against background noise and more compact with contextual information, can further improve the performance of the DNN baseline. Compared with the standard Gaussian Mixture Model (GMM) baseline of the DCASE 2016 audio tagging challenge, our proposed method obtains a significant equal error rate (EER) reduction from 0.21 to 0.13 on the development set. The proposed aDAE system can get a relative 6.7% EER reduction compared with the strong DNN baseline on the development set. Finally, the results also show that our approach obtains the state-of-the-art performance with 0.15 EER on the evaluation set of the DCASE 2016 audio tagging task while EER of the first prize of this challenge is 0.17.

Yong Xu, Qiang Huang, Wenwu Wang et al.

Acoustic event detection for content analysis in most cases relies on lots of labeled data. However, manually annotating data is a time-consuming task, which thus makes few annotated resources available so far. Unlike audio event detection, automatic audio tagging, a multi-label acoustic event classification task, only relies on weakly labeled data. This is highly desirable to some practical applications using audio analysis. In this paper we propose to use a fully deep neural network (DNN) framework to handle the multi-label classification task in a regression way. Considering that only chunk-level rather than frame-level labels are available, the whole or almost whole frames of the chunk were fed into the DNN to perform a multi-label regression for the expected tags. The fully DNN, which is regarded as an encoding function, can well map the audio features sequence to a multi-tag vector. A deep pyramid structure was also designed to extract more robust high-level features related to the target tags. Further improved methods were adopted, such as the Dropout and background noise aware training, to enhance its generalization capability for new audio recordings in mismatched environments. Compared with the conventional Gaussian Mixture Model (GMM) and support vector machine (SVM) methods, the proposed fully DNN-based method could well utilize the long-term temporal information with the whole chunk as the input. The results show that our approach obtained a 15% relative improvement compared with the official GMM-based method of DCASE 2016 challenge.