@techreport{Bando_2025_t4,
    Author = "Nozaki, Yuto and Sakurai, Shun and Bando, Yoshiaki and Saijo, Kohei and Imoto, Keisuke and Onishi, Masaki",
    title = "A HYBRID S5 SYSTEM BASED ON NEURAL BLIND SOURCE SEPARATION",
    institution = "DCASE2025 Challenge",
    year = "2025",
    month = "June",
    abstract = "In this paper, we report our hybrid system for the DCASE 2025 Challenge Task 4 based on neural blind source separation (BSS). This task, called spatial semantic segmentation of sound scenes (S5), aims to detect and separate sound events from a multichannel mixture signal. To make the separation model robust against unseen audio environments, we leverage neural BSS to combine robust statistical signal processing and high-performing neural modeling. Specifically, our network architecture incorporates the iterative source steering algorithm to separate source signals using spatial statistics. The network is trained via multitask learning of source separation and classification with permutation invariant training. In addition, to improve the performance, we utilized an audio foundation model called BEATs and augmented the training data by curating AudioSet. The experimental results on the official development test set show that our best system (System 2) improved more than 2 dB in class-aware signal-to-distortion ratio improvement (CA-SDRi) from the official baseline system."
}

@techreport{Choi_2025_t4,
    Author = "Kwon, Younghoo and Lee, Dongheon and Kim, Dohwan and Choi, Jung-Woo",
    title = "SELF-GUIDED TARGET SOUND EXTRACTION AND CLASSIFICATION THROUGH UNIVERSAL SOUND SEPARATION MODEL AND MULTIPLE CLUES",
    institution = "DCASE2025 Challenge",
    year = "2025",
    month = "June",
    abstract = "This paper presents a multi-stage framework that integrates Universal Sound Separation (USS) and Target Sound Extraction (TSE) for sound separation and classification. In the first stage, USS is applied to decompose the input audio into multiple source components. Each separated source is then individually classified to generate two types of clues: enrollment and class clues. These clues are subsequently utilized in the second stage to guide the TSE process. By generating multiple clues in a self-guided manner, the proposed method enhances the performance of target sound extraction. The final output of the TSE module is then re-classified to improve the classification accuracy further."
}

@techreport{FulinWu_2025_t4,
    Author = "Wu, Fulin and Wang, Zhong-Qiu",
    title = "TS-TFGRIDNET: EXTENDING TFGRIDNET FOR LABEL-QUERIED TARGET SOUND EXTRACTION VIA EMBEDDING CONCATENTAITON",
    institution = "DCASE2025 Challenge",
    year = "2025",
    month = "June",
    abstract = "The DCASE2025 Challenge Task 4 - Spatial Semantic Segmentation of Sound Scenes (S5) challenges participants to separate a set of mixed sound events (sampled from 18 targeted sound events) to individual sound-event signals. The baseline system provided by the challenge organizors first performs audio tagging to identify the sound events existed in the mixture, and then conducts label-queried target sound extraction (TSE) to extract the signal of each identified sound event. Building on the baseline system, we propose to improve the label-queried TSE component by using a novel model named Target Sound Extraction TF-GridNet (TS-TFGridNet), leveraging the strong capability of TF-GridNet at speech separation for TSE. TS-TFGridNet concatenates audio and sound-class embeddings along the frequency or feature dimension, thereby conditioning TF-GridNet to perform TSE. Clear improvement is observed over the baseline system."
}

@techreport{Morocutti_2025_t4,
    Author = "Morocutti, Tobias and Schmid, Florian and Greif, Jonathan and Primus, Paul and Widmer, Gerhard",
    title = "TRANSFORMER-AIDED AUDIO SOURCE SEPARATION WITH TEMPORAL GUIDANCE AND ITERATIVE REFINEMENT",
    institution = "DCASE2025 Challenge",
    year = "2025",
    month = "June",
    abstract = "This technical report presents the CP-JKU team’s system for Task 4 Spatial Semantic Segmentation of Sound Scenes of the DCASE 2025 Challenge. Building on the two-stage baseline of audio tagging followed by label-conditioned source separation, we introduce several key enhancements. We reframe the tagging stage as a sound event detection task using five Transformers pre-trained on AudioSet Strong, enabling temporally strong conditioning of the separator. We further inject the Transformer’s latent representations into a ResUNet separator initialized from AudioSep and extended with a Dual-Path RNN. Additionally, we propose an iterative refinement scheme that reuses the separator’s output as input. These improvements raise label prediction accuracy to 73.07\% and CASDRi to 14.49 for a single-model system on the development test set, substantially surpassing the baseline."
}

@techreport{Park_2025_t4,
    Author = "Park, Jongyeon and Lee, Joonhee and Lim, Do-Hyeon and Kim, Hong Kook and Geum, Hyeongcheol and Lim, Jeong Eun",
    title = "PERFORMANCE IMPROVEMENT OF SPATIAL SEMANTIC SEGMENTATION WITH ENRICHED AUDIO FEATURES AND AGENT-BASED ERROR CORRECTION FOR DCASE 2025 CHALLENGE TASK 4",
    institution = "DCASE2025 Challenge",
    year = "2025",
    month = "June",
    abstract = "This technical report presents submission systems for Task 4 of the DCASE 2025 Challenge. This model incorporates additional audio features (spectral roll-off and chroma features) into the embedding feature extracted from the mel-spectral feature to improve the classification capabilities of an audio-tagging model in the spatial semantic segmentation of sound scenes (S5) system. This approach is motivated by the fact that mixed audio often contains subtle cues that are difficult to capture with mel-spectrograms alone. Thus, these additional features offer alternative perspectives for the model. Second, an agent-based label correction system is applied to the outputs processed by the S5 system. This system reduces false positives, improving the final class-aware signal-to-distortion ratio improvement (CA-SDRi) metric. Finally, we refine the training dataset to enhance the classification accuracy of low-performing classes by removing irrelevant samples and incorporating external data. That is, audio mixtures are generated from a limited number of data points; thus, even a small number of out-of-class data points could degrade model performance. The experiments demonstrate that the submitted systems employing these approaches relatively improve CA-SDRi by up to 14.7\% compared to the baseline of DCASE 2025 Challenge Task 4."
}

@techreport{Qian_2025_t4,
    Author = "Zhou, Xin and Wang, Hongyu and Li, Chenda and Han, Bing and Zheng, Xinhu and Qian, Yanmin",
    title = "SJTU-AUDIOCC SYSTEM FOR DCASE 2025 CHALLENGE TASK 4: SPATIAL SEMANTIC SEGMENTATION OF SOUND SCENES",
    institution = "DCASE2025 Challenge",
    year = "2025",
    month = "June",
    abstract = "The present report introduces four systems developed by the AudioCC Lab at Shanghai Jiao Tong University for DCASE 2025 Task 4. The task at hand is to detect target sound events and separate corresponding signals from multi-channel mixture. It was found that the effective detection of sound events and extraction of the corresponding signals was challenging under conditions where mixture consists of multiple target sound events, non-target sound events, and non-directional background noise. In order to address this challenge, we propose four systems. The first system represents an enhancement to the baseline system, the second is a multistage iterative system that is both novel and promising, the third is a lightweight model based on Encoder-Decoder Attractor (EDA) module, and the fourth integrates multiple audio tagging models to achieve optimal performance. These four systems cover high performance, low overhead, and promising frameworks, providing a reference for future research on this task."
}

@techreport{Stergioulis_2025_t4,
    Author = "Stergioulis, Vasileios and Potamianos, Gerasimos",
    title = "REDUX: AN ITERATIVE STRATEGY FOR SEMANTIC SOURCE SEPARATION",
    institution = "DCASE2025 Challenge",
    year = "2025",
    month = "June",
    abstract = "In this report, we present a system for the Spatial Semantic Segmentation of Sound Scenes (DCASE 2025 Task 4), combining enhanced source separation and label classification through an iterative verification strategy. Our approach integrates the Masked Modeling Duo (M2D) classifier with a separator architecture based on an attentive ResUNeXt network. Inspired by recent advances in universal audio modeling and self-supervised separation, our system incorporates feedback between multiple classification and separation stages to correct early-stage prediction errors. Specifically, classification outputs are verified using post-separation reclassification, and ambiguous cases are resolved through targeted waveform subtraction and re-analysis. This strategy enables improved source-label associations without increasing model complexity. Evaluated on the development set, our method achieves a relative improvement of 0.28\% in CA-SDRi and 1.46\% in accuracy over the baseline."
}

@techreport{Zhang_2025_t4,
    Author = "Wang, Zehao and Wang, Sen and Zhang, Zhicheng and Yin, Jianqin",
    title = "SPATIAL SEMANTIC SEGMENTATION OF SOUND SCENES BASED ON ADAPTER FINE-TUNING",
    institution = "DCASE2025 Challenge",
    year = "2025",
    month = "June",
    abstract = "In this work, we present our submission system for DCASE 2025 Task4 on Spatial semantic segmentation of sound scenes (S5).Among them, we introduce the audio tagging (AT) and labelquery source separation (LSS) systems built on the fine-tuned M2D and the modified version of ResUnet. By introducing the bidirectional recurrent neural network (DPRNN) module into ResUNet and improving the Feature-wise Linear Modulation (FiLM) mechanism, the model’s ability to capture long-term dependent features in spatial audio and the flexibility of dynamic feature adjustment are enhanced. Experimental results show that the improved system outperforms the baseline system in class-aware evaluation metrics (CA-SDRi, CA-SI-SDRi), verifying the effectiveness of the method."
}