@techreport{Bando_AIST2026,
    Author = "Bando, Yoshiaki and Sakurai, Shun and Nozaki, Yuto and Imoto, Keisuke and Onishi, Masaki",
    title = "END-TO-END ITERATIVE S5 SYSTEM BASED ON TF-LOCOFORMER AND ATST-FRAME",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report describes our end-to-end (E2E) system based on TF-Locoformer and ATST-Frame. The best system in the previous challenge on spatial semantic segmentation of sound scenes (S5) leveraged an iterative architecture that alternately performs separation and classification to achieve excellent performance. Inspired by this architecture, we built an E2E system that iteratively applies TF-Locoformer and ATST-Frame. Specifically, the overall architecture is based on TF-Locoformer, which stacks Transformerbased blocks to process each time-frequency bin. We inserted signal and classification heads into the outputs of intermediate blocks, and applied ATST-Frame to the separated source signals. The sourcewise embeddings extracted by ATST-Frame are then fed back to the Locoformer block to progressively improve the performance. The whole architecture is trained in an E2E manner with permutation invariant training. Our best model on the development set achieved a class-aware permutation invariant signal-to-distortion ratio improvement (CAPI-SDRi) of 16.3 dB and source-wise label accuracy of 73.6 \% on the test subset of the development set."
}

@techreport{Choi_KAIST2026,
    Author = "Kwon, Younghoo and Park, Junwoo and Yin, Han and Choi, Jung-Woo",
    title = "A MULTI-STAGE SEPARATION-AND-CLASSIFICATION FRAMEWORK GUIDED BY COMPLEMENTARY ACOUSTIC-TO-SEMANTIC CLUES",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This report describes the system proposed for the DCASE 2026 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes (S5). Specifically, we develop a multi-stage framework in which each stage couples a separation model with a classification model. The first stage performs source separation and classification directly on the multi-channel mixture. Its outputs are then propagated to the following stage as two complementary clues that progressively refine each target estimate: (i) an enrollment clue, the separated waveform itself, serving as a low-level acoustic reference; and (ii) a class clue, the predicted label encoded as a one-hot vector. The third stage reuses the second-stage outputs under the same scheme, forming an iterative self-guided refinement process. In addition, we use a fine-grained frame-level audio embedding from an audio encoder pretrained on a large audio corpus as an additional clue to further improve the audio separation performance. On the test set, the proposed system achieves a CAPI-SDRi of 15.51 dB, a mixture accuracy of 71.09\%, and a source accuracy of 78.62\%; with an improvement of 7.02 dB, 10.38\%p and 8.22\%p compared with the challenge baseline, respectively."
}

@techreport{Deng_WHU2026,
    Author = "Deng, Yongyi and Zou, Tong and Tian, Yanxin and Shi, Hao and Luo, Jiayue and Yan, Yicheng and Huang, Gongping",
    title = "A LABEL-GUIDED ENSEMBLE SYSTEM FOR SPATIAL SEMANTIC SEGMENTATION OF SAME-CLASS SOUND SOURCES",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This report presents our system for DCASE 2026 Task 4, which addresses spatial semantic segmentation of sound scenes containing same-class foreground sources and inactive source labels. The task requires not only separating active sound events that may share the same class label, but also handling source slots corresponding to inactive labels. To improve the reliability of label-conditioned source separation, we adopt a label-guided ensemble strategy. In the tagging stage, two M2D-AT variants and one CLAPAT variant are fused by weighted voting to obtain robust source-label estimates. The estimated labels are then used to guide source separation. The primary separator is a fine-tuned ResUNetK model with a mask-sharpen inference variant, while a TF-GridNet model is used only as a weak auxiliary branch for a small number of selected classes through class-dependent fusion weights. Instead of uniformly averaging separator outputs, the final outputs are generated through label-guided class-dependent fusion, which improves the consistency between predicted labels and separated sources while keeping inactive slots controlled by silence-label conditioning. On the development set, according to the submission-pack evaluation snapshot, our final system achieves a CAPI-SDRi of 8.625, with a mixture-level accuracy of 61.706\% and a source-level accuracy of 72.139\%."
}

@techreport{Jeong_Medisensing2026,
    Author = "Jeong, Seunggyu and Kim, Seong-Eun",
    title = "TAGGING-DRIVEN INFERENCE REFINEMENT AND DUAL-SEPARATOR SELECTION FOR SPATIAL SEMANTIC SEGMENTATION OF SOUND SCENES",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "We describe our submission to DCASE 2026 Challenge Task 4, spatial semantic segmentation of sound scenes. A system has to detect which sound classes are active in a four-channel mixture and to separate each detected source, and is scored by class-aware permutation-invariant signal-to-distortion ratio improvement (CASDRi). We build on the official baseline, which couples two Masked Modeling Duo (M2D) audio taggers with a FiLM-conditioned ResUNet label-queried separator. An oracle study, in which groundtruth labels are fed to the unchanged baseline separator, reaches 9.52 dB, so for the given separator the label decision is the limiting factor up to about 9.5 dB. Our system follows this finding and concentrates on the labelling and decision stages, almost all of it at inference time. We fuse the single- and four-channel taggers, fine-tune the four-channel tagger with a curriculum that oversamples silent-target clips, re-tag the separated stems to verify and clean the queries, gate residual false positives using the reward structure of CA-SDRi, and select stems from two separators by re-tagging agreement. These steps raise development-set CA-SDRi from 8.49 dB to 9.06 dB and mixture tagging accuracy from 60.7\% to 63.8\%. We submit three systems that add these components in turn."
}

@techreport{Park_KUBIG2026,
    Author = "Park, Yoohan and Cho, Haejin",
    title = "END-TO-END SPATIAL SEMANTIC SEPARATOR WITH DOA MODULE",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report describes our submission to DCASE 2026 Challenge Task 4, Spatial Semantic Segmentation of Sound Scenes. The task requires a system to separate sound events from 4-channel First-Order Ambisonics (FOA) recordings while simultaneously predicting their sound classes and directions of arrival (DoA). To address this challenge, we propose an end-to-end framework that jointly performs source separation, sound classification, and DoA estimation. The model generates a fixed set of source representations, each associated with a separated waveform, a class label, and a DoA estimate. A frozen BEATs encoder is used to provide robust acoustic representations, while lightweight task-specific modules are trained for spatial modeling and prediction. The entire system is optimized using a joint permutation-invariant training objective that encourages consistent source assignment across all outputs."
}

@techreport{Park_SGU2026,
    Author = "Ku, Bon-Hyeok and Jeong, Woocheol and Park, Hyung-Min",
    title = "EXTENDING SR-CORRNET TO LABEL-QUERIED TARGET SOUND EXTRACTION",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This paper describes our submission to DCASE 2026 Task 4. We extend SR-CorrNet, originally designed for blind source separation, into a label-queried target sound extraction and separation model. The system reformulates blind separation into target extraction by conditioning the separator with frame-level strong class labels via Feature-wise Linear Modulation (FiLM), steering each output slot to extract the queried class. To stabilize ambiguous regions, we supplement this with a time-invariant weak class label FiLM bias. The class labels are predicted by a front-end fusion tagger that combines two complementary AudioSet-pretrained Transformers (M2D and PaSST) via feature-axis concatenation. The extended model operates with block streaming inference, coupling the tagger and separator through a soft-query interface."
}

@techreport{Saijo_Mitsubishi2026,
    Author = "Saijo, Kohei and Masuyama, Yoshiki and Boeddeker, Christoph and Richter, Julius and Edo, Takahiro and Wichern, Gordon and Roux, Jonathan Le",
    title = "THE MERL SYSTEMS FOR DCASE 2026 CHALLENGE TASK 4",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report describes our spatial semantic segmentation of sound scenes (S5) systems for DCASE 2026 Challenge Task 4. Inspired by the top-ranked system in DCASE 2025 Task 4, we adopt a cascaded framework consisting of universal sound separation (USS) with source counting, source classification, and class-aware refinement. In the first stage, a TF-Locoformer-based USS model separates multi-channel mixtures into single-channel foreground and interference signals. Then, each separated signal is classified into one of 18 foreground classes or as interference. The separated foreground signals are further refined by another TF-Locoformer-based model conditioned on the predicted class labels and the observed mixture. Our best system achieves CA-PI-SDRi of 14.95 dB and mixture accuracy of 78.11\% on the dev test set."
}

@techreport{Wang_BUPT2026,
    Author = "Wang, Sen and Tang, Chengyao and Zhang, Zhicheng and Yin, Jianqin",
    title = "SEMANTIC DISTILLATION FOR SPATIAL SEMANTIC SEGMENTATION OF SOUND SCENES",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This report describes our system for DCASE 2026 Task 4, spatial semantic segmentation of sound scenes. The system follows a twostage pipeline: an M2D-based audio tagger first predicts up to three event labels, and a label-queried ResUNetK then separates the corresponding dry monaural sources from a four-channel spatial mixture. We improve audio tagging along two complementary directions. The first uses permutation-invariant deep supervision and an exponential-moving-average teacher. The second transfers semantic representations from a frozen Qwen2-Audio model to M2D using centered-kernel-alignment and cosine objectives. Both methods are developed for single-channel (1c) and four-channel (4c) tagging. On the development test set, the submitted 1c system obtains 8.557 dB CAPI-SDRi, while the submitted 4c system obtains 8.807 dB CAPI-SDRi with 61.442\% mixture-level label accuracy and 72.535\% source-level label accuracy."
}

@techreport{Wang_SRCN2026,
    Author = "Wang, Ruohan and Chen, Minjun and Liu, Yangyang and Wu, Longhai and Chen, Jie",
    title = "Local-Global Transformer with Iterative Refinement for Multi-Channel Sound Source Separation and Extraction",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report describes our proposed systems for DCASE2026 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes (S5). The task aims to enhance technologies for sound event detection and separation from multi-channel input signals that mix multiple sound events with spatial information, which is a fundamental basis of immersive communication. Our approach employs a deep frequency-time transformer architecture with local modeling by convolution that processes multi-channel audio recordings from a 4-microphone array. The system consists of three main components: (1) a universal sound separation module for separating waveform and predicting the sound classes at the same time, (2) an audio tagging model for semantic label prediction, and (3) an iterative target sound extraction module that leverages enrollment clues and semantic labels to extract specific sound sources. We incorporate spatial features including interchannel phase difference (IPD) and inter-channel level difference (ILD) to enhance separation performance."
}

@techreport{You_PKU2026,
    Author = "You, Yuhuan",
    title = "DCASE 2026 Task 4 Submission: Duplicate-Label-Aware Source Selection and TUSS Overlay Grafting",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This report describes the PKU submission to DCASE 2026 Challenge Task 4, Spatial Semantic Segmentation of Sound Scenes. The task requires detection and separation of target sound events from FOA mixtures, and the 2026 setting includes same-class multiple sources and zero-target soundscapes. Our submitted systems use the official baseline family as the separation and tagging foundation, then add duplicate-label-aware candidate selection and label-conditioned overlay grafting. The main ranking metric is CAPI-SDRi, so the system is optimized to select separated sources that jointly improve class assignment and permutation-invariant signal quality. We submit four systems: a stable duplicate-label weighted selector, two progressively more aggressive TUSS overlay systems, and a full all-label TUSS candidate-pool selector for diversity."
}