Lin Wang

Speech enhancement in the presence of strong drone noise at extremely low signal-to-noise ratios (SNRs) remains challenging, and conventional objective metrics may fail to accurately reflect perceived speech quality. In this paper, we evaluate several deep learning-based speech enhancement algorithms under severe drone-noise conditions using a diverse set of objective metrics, including signal fidelity, intelligibility, and perceptual quality measures. The results reveal substantial inconsistencies among metrics, leading to conflicting conclusions regarding algorithm performance. To address this issue, we conduct a controlled subjective listening experiment under extremely low-SNR conditions to assess the perceptual relevance of these metrics. The consistency analysis shows that learning-based perceptual metrics (e.g., NISQA) and intelligibility-oriented measures (e.g., ESTOI) achieve the highest agreement with subjective judgments, followed by SI-SDR, while traditional quality metrics such as PESQ exhibit weaker agreement. These findings highlight the limitations of conventional metrics in extremely low-SNR conditions and provide guidance for selecting perceptually meaningful evaluation criteria in drone audition.

Bimodal speech emotion recognition is challenging due to the difficulty of capturing subtle emotional cues and effectively modeling complex interactions between audio and text. Existing methods often overlook the importance of modality connections and struggle to filter noisy or misleading cross-modal signals. To address these issues, we propose a novel Bimodal Connection Attention Fusion (BCAF) framework that explicitly models modality-specific representations and their interactions. BCAF includes three key modules: an interactive connection network that learns modality connections through an encoderdecoder structure; a bimodal attention network that captures intra- and inter-modal dependencies; and a correlative attention network that filters cross-modal noise to improve robustness. Extensive evaluations across benchmark datasets show that BCAF achieves superior performance over existing methods, highlighting its strength in modeling detailed modality connections and interactions for emotion recognition.

Aerial bioacoustic monitoring has the potential to transform biodiversity surveys by enabling rapid acoustic sampling across large and inaccessible habitats. However, its application from unmanned aerial vehicles (UAVs) has remained limited because strong rotor self‐noise overwhelms target sounds such as bird vocalisations, compromising the reliability of ecological information derived from onboard recordings. We present a practical drone‐based acoustic sensing system that combines targeted hardware noise mitigation with a deep‐learning noise reduction workflow to recover bird vocal activity from highly noisy UAV recordings. The system incorporates a locally adaptive enhancement model trained on in situ noise characteristics, enabling robust deployment under real field conditions. Using in situ UAV recordings, bird vocalisations were recovered at flight altitudes of up to 30 m, despite input signal‐to‐noise ratios typically ranging from −20 to −30 dB. Noise suppression substantially improved automated bird detection rates using off‐the‐shelf species recognition software, and community similarity analyses showed that denoised aerial recordings closely matched ground‐based and ambient reference soundscapes. By restoring the ecological interpretability of drone‐based acoustic data, this work demonstrates that aerial bioacoustic monitoring can function as a practical and scalable tool for biodiversity assessment in challenging environments, enabling surveys in habitats that are otherwise difficult or impossible to access.

Global efforts to mitigate anthropogenic pressures on biodiversity and ecosystems will often be realised through management at landscape-scales (i.e., in the range of 100s–1000 s km²). In consequence, we need to measure biodiversity responses at landscape-scales to ensure mitigations are effectively protecting and restoring ecosystems. Yet many countries currently lack monitoring programmes that can generate indicators of biodiversity at these scales. Localised monitoring (e.g., 1 km²) is often amalgamated into national-scale indicators, however, this leaves a substantial gap in the middle of this spatial gradient, limiting availability of information at decision-relevant scales. Here, using the United Kingdom as a case study, we explored the suitability of seven sources of biodiversity data which could be used to construct landscape-scale indicators. We surveyed 70, mostly UK-based, monitoring experts for their opinions on structured and unstructured in-person surveys, camera traps, eDNA, drones, passive acoustic recorders, and satellite remote sensing. We assessed data source utility to construct indicators reflecting Essential Biodiversity Variables, i.e., as holistic measures of taxa or ecosystems rather than assessments of individual management interventions. All seven data sources were deemed suitable, and experts expected developments in technology and infrastructure to greatly increase this potential over the next decade. However, there are technical, analytical, logistical and financial barriers to establishing monitoring networks that could yield the requisite data for landscape-scale indicators. Resolving these issues requires substantial research, policy commitment and investment, but landscape-scale indicators will be essential for the UK to undertake adaptive management and monitor nature recovery.

Linear Text Segmentation is the task of automatically tagging text documents with topic shifts, i.e. the places in the text where the topics change. A well-established area of research in Natural Language Processing, drawing from well-understood concepts in linguistic and computational linguistic research, the field has recently seen a lot of interest as a result of the surge of text, video, and audio available on the web, which in turn require ways of summarising and categorizing the mole of content for which linear text segmentation is a fundamental step. In this survey, we provide an extensive overview of current advances in linear text segmentation, describing the state of the art in terms of resources and approaches for the task. Finally, we highlight the limitations of available resources and of the task itself, while indicating ways forward based on the most recent literature and under-explored research directions.

The paper summarizes the contributions of participants to the sixth Sussex-Huawei Locomotion-Transportation (SHL) Recognition Challenge organized at the HASCA Workshop of UbiComp/ISWC 2024. The goal of this machine learning/data science challenge is to recognize eight locomotion and transportation activities (Still, Walk, Run, Bike, Bus, Car, Train, Subway) from the motion (accelerometer, gyroscope, magnetometer) sensor data of a smartphone in a way which is user-independent and smartphone position-independent, and as well robust to data missing during deployment. The training data of a 'train' user is available from smartphones placed at four body positions (Hand, Torso, Bag and Hips). The testing data originates from 'test' users with a smartphone placed at one of three body positions (Torso, Bag or Hips). In addition, the test data has one or multiple sensor modalities randomly missing from each time frame (5 seconds). Such a scenario may occur if a device turns on and off dynamically sensors to save power, or due to limited computational or memory capacity. We introduce the dataset used in the challenge and the protocol of the competition. We present a meta-analysis of the contributions from 7 submissions, their approaches, the software tools used, computational cost and the achieved results. Overall, one submission achieved an F1 score between 70% and 80%, two between 60% and 70%, three between 50% and 60%, and one below 50%. Finally, we present a baseline implementation addressing missing sensor modalities.

Iacopo Ghinassi, Lin Wang, Chris Newell, Matthew Purver. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024.

Neural sentence encoders (NSE) are effective in many NLP tasks, including topic segmentation. However, no systematic comparison of their performance in topic segmentation has been performed. Here, we present such a comparison, using supervised and unsupervised segmentation models based on NSEs. We first compare results with baselines, showing that the use of NSEs does often provide improvements, except for specific domains such as news shows. We then compare over three different datasets a range of existing NSEs and a new NSE based on ad hoc pre-training strategy. We show that existing literature documenting general performance gains of NSEs does not always conform to the results obtained by the same NSEs in topic segmentation. If Transformers-based encoders do improve over previous approaches, fine-tuning in sentence similarity tasks or even on the same topic segmentation task we aim to solve does not always equate to better performance, as results vary across method being used and domains of application. We aim to explain this phenomenon and the relative poor performance of NSEs in news shows by considering how well different NSEs encode the underlying lexical cohesion of same-topic segments; to do so, we introduce a new metric, ARP. The results from this study suggest that good topic segmentation results do not always rely on good cohesion modelling on behalf of the segmenter and that is dependent upon what kind of text we are trying to segment. Also, it appears evident that traditional sentence encoders fail to create topically cohesive clusters of segments when used on conversational data. Overall, this work advances our understanding of the use of NSEs in topic segmentation and of the general factors determining the success (or failure) of a topic segmentation system. The new proposed metric can quantify the lexical cohesion of a multi-topic document under different sentence encoders and, as such, might have many different uses in future research, some of which we suggest in our conclusions.