We introduce the REAL-TSE Challenge, an IEEE SLT 2026 satellite challenge on target speaker extraction~(TSE) from real conversational recordings. Given a multi-speaker mixture and one or more enrollment utterances from a target speaker, participating systems must recover only the target speech. Unlike simulated read-speech benchmarks, REAL-TSE evaluates Mandarin and English recordings that contain natural overlap, reverberation, noise, channel mismatch, and conversational dynamics. The challenge defines two complementary tracks: an Online track for low-latency streaming extraction and an Offline track for full-context processing. Systems are evaluated with Token Error Rate (TER), Speaker Similarity (SpkSim), DNSMOS, and target-speaker activity F1. This overview paper describes the task definition, datasets, baselines, evaluation protocol, submitted systems, condition-wise findings, and lessons for future real-world TSE benchmarks.
Shuai Wang, Zihan Qian, Ke Zhang et al.· arXiv.org· 1 citation
In complex and unstructured orchard environments, accurate fruit detection is essential for yield estimation and robotic harvesting in precision agriculture. However, single-model detectors often suffer from reduced robustness and high miss rates under drastic illumination changes, severe occlusions, and dense fruit overlap. To address these challenges, this study proposes a multi-model fusion framework for robust mango detection in complex orchard environments. The proposed method employs YOLOv8n, YOLOv8s, and YOLOv8m as base detectors and applies multi-scale test-time augmentation (TTA) to obtain predictions from different augmented views. After mapping the predicted bounding boxes back to the original image coordinate system, predictions corresponding to the same target across different TTA views of each base detector are matched based on the intersection over union (IoU), yielding model-specific prediction results. Weighted Box Fusion (WBF) is then applied to determine the fused bounding-box coordinates. For candidate targets jointly detected by multiple base detectors, the confidence scores provided by the individual models are combined using Noisy-OR to obtain the fused confidence score. Finally, Gaussian Soft-NMS is applied to decay the scores of overlapping candidate boxes, thereby reducing the risk of incorrectly suppressing adjacent mangoes in densely clustered scenes. Experiments on two complementary datasets under within-dataset evaluation protocols demonstrate the effectiveness of the proposed method. On the standard dataset (Data1), Recall and mAP@0.5 reach 95.52% and 98.60%, respectively. Across five repeated random holdout splits of Data2, the proposed framework increased the mean Recall from 82.79% to 84.90% and the mean mAP@0.5 from 90.27% to 91.23%. These results indicate that the proposed framework improves detection robustness and completeness compared with single-model detectors in complex orchard environments, demonstrating its potential for offline yield estimation and orchard phenotyping.
