@techreport{Colotti_TAU_2026,
    Author = "Colotti, Francesco and Markl, Kerstin and Casciotti, Riccardo and Naranjo, Javier",
    title = "ACACIA: Audio Classification using Attention-Based Cleaned Multimodal Embeddings",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This paper presents the ACACIA pipeline, submitted to Task 1 of DCASE 2026, the first edition of the Heterogeneous Audio Classification challenge based on the Broad Sound Taxonomy (BST). Our approach combines complementary acoustic and textual information available in Freesound recordings by jointly exploiting the audio signal together with user-provided tags and textual descriptions. To improve the quality of the textual modality, we propose a preprocessing and cleaning pipeline that removes noisy and non-informative content from descriptions before encoding. The proposed architecture comprises three dedicated branches that independently encode audio, tags, and descriptions, whose representations are fused into a shared multimodal embedding space. To exploit the hierarchical organization of BST, the model performs two classification tasks simultaneously, predicting both top-level and second-level categories, and is trained using the sum of two binary cross-entropy losses. Experimental results show that ACACIA surpasses the official baseline by 6\\% according to the challenge's hierarchy-aware evaluation metric, demonstrating that the integration of acoustic information with complementary textual metadata provides a robust and effective solution for heterogeneous sound classification in real-world conditions."
}

@techreport{Guan_HEU_2026,
    Author = "Xiaoyu, Feng and Tong, Ye and Xuefeng, Yang and Feiyang, Xiao and Zhu, Qiaoxi and Guan, Jian",
    title = "GISP@HEU's Submission for Task 1: Heterogeneous Audio Classification in the DCASE 2026 Challenge",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report presents our submitted systems for Task 1: Heterogeneous Audio Classification in the Detection and Classification of Acoustic Scenes and Events (DCASE) 2026 Challenge. Our submission consists of four systems, including three individual systems and one ensemble system. System 1 and System 2 adopt a two-stage hierarchical training framework with hyperbolic representation learning, and are built upon different multimodal audio language backbones, i.e., Qwen2-Audio and Qwen3-Omni. System 3 is developed from the official baseline framework, using multimodal embeddings and hierarchical training with clean BSD10k data and filtered BSD35k data. System 4 is an ensemble system consisting of these three systems. Experimental results demonstrate that the proposed approach improves hierarchical classification performance and achieves 81.80\\% ± 0.21\\% on the development dataset."
}

@techreport{Han_CSU_2026,
    Author = "Han, Sarang and Jo, Minsik and Ha, Eunseo and Chae, Minju and Kang, Hyeonguk and Lee, Geonwoo",
    title = "MESH:MULTI-EMBEDDINGSYSTEMWITHHIERARCHICALPROXYLEARNINGFOR DCASE 2026 TASK 1",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report describes a multi-embedding system with hierarchical proxy learning (MESH) for DCASE 2026 Challenge Task 1. The task is based on the Broad Sound Taxonomy (BST) for hierarchical audio classification, where each sample is assigned to a lower-level category within a top-level class. Fixed audio embeddings were extracted from multiple pretrained models and pooling configurations, and the final embedding set was selected using validation hierarchical F-score (hF) and embedding diversity. The selected audio embeddings were concatenated with fixed CLAP text embeddings, and the classifier was trained through BSD35k-CS pre-training followed by BSD10k-v1.2 fine-tuning. To incorporate the BST hierarchy, a hierarchical proxy loss included distinct proxy sets for top-level and lower-level classes. Final predictions were aggregated with selective kernel-based output fusion and OOF-based ensemble selection, and the best OOF ensemble achieved 84.59\% hF among the four submitted systems."
}

@techreport{Huang_WHU_2026,
    Author = "Ning, Beile and Yu, Jiayi and Wang, Zitong and Hu, Yufei and Xu, Wenjun and Qian, Yuanhang and Bai, Zhongxin and Huang, Gongping",
    title = "A MULTI-BRANCH AND HIERARCHY-AWARE SYSTEM FOR BST AUDIO CLASSIFICATION",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report introduces our system for Task 1 of the DCASE 2026 Challenge. Our system leverages multimodal audio-text representations extracted by CLAP and further improves classification performance through dataset expansion, branch-enhanced acoustic modeling, and hierarchy-aware prediction strategies. To improve data diversity, we selected a subset of the BSD35K dataset and incorporated it into BSD10K to expand the training set. We introduced additional acoustic branches based on Mel-Frequency Cepstral Coefficients (MFCC), log-Mel spectrogram, and log Short-Time Fourier Transform (log-STFT), and adopted different hierarchical strategies and post-processing strategies, with the post-processing strategy further integrated into the system via knowledge distillation. During training, we also applied data augmentation techniques including time masking and random cropping. Our best single system extracts log-STFT features from audio to facilitate model training on the new dataset, and employs post-processing to refine model predictions, achieving a hierarchical F1 score of 80.84\% on the BSD10K test set under the same evaluation protocol as the baseline. Furthermore, we constructed two ensemble models based on different backbone architectures, obtaining hierarchical F1 scores of 81.25\% and 81.18\%, respectively, on the same BSD10K test set as the baseline."
}

@techreport{Kil_Medisensing_2026,
    Author = "Kil, Minkyu and Jeong, Seunggyu and Kim, Seong-Eun",
    title = "POSTERIOR STACKING AND CONSERVATIVE METADATA GATING FOR HETEROGENEOUS AUDIO CLASSIFICATION",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report describes the Medisensing-SeoulTech submission to DCASE 2026 Task 1, Heterogeneous Audio Classification. The submitted systems classify each evaluation item into one of 23 second-level Broad Sound Taxonomy classes. The final package contains four systems. Systems 1 and 2 are high-scoring weighted posterior stacks built from frozen audio and audio-text embedding heads with conservative same-parent metadata gates. System 3 is a Larger-CLAP residual classifier with a TF-IDF metadata gate, included as a complementary metadata-aware variant with broader label-symbol coverage. System 4 is a complementary audio-only posterior ensemble. Across metadata-assisted systems, metadata can refine the second-level class only inside the audio-predicted top-level parent. No evaluation-set ground-truth labels or manual evaluation-set annotations are used for training, threshold selection, system selection, or reporting."
}

@techreport{Kucukoglu_NYU_2026,
    Author = "Kucukoglu, Mehmet Atilay",
    title = "MULTIMODAL ENSEMBLE SYSTEM FOR HETEROGENEOUS AUDIO CLASSIFICATION",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report details an ensemble approach and experiments leading to it as a submission to DCASE 2026 Challenge Task 1 on Heterogeneous Audio Classification using the Broad Sound Taxonomy (BST). The official HATR baseline system achieves 79.01\% hierarchical F1 score on BSD10k-v1.2. Alternative audio encoders such as BEATs, ConvNeXt, MATPAC, and Whisper, performed below CLAP, suggesting that CLAP’s joint audio-text alignment is a powerful approach for this task. Incorporating hierarchical loss, contrastive loss, confidence weighting, and class weighting did not improve the hierarchical F1 score. Among data augmentation methods, cross-modal swap with Gaussian noise achieved the best single model result at 79.13\% hF1. Incorporating external data from FSD50K and ESC-50 mapped to the BST taxonomy did not achieve higher F1 scores. End-to-end fine-tuning of the CLAP audio encoder was also investigated but did not surpass the frozen multimodal baseline. The best result was obtained using an ensemble of 5 models incorporating different encoders, loss strategies, and augmentation methods, achieving 81.13\% hierarchical F1 score on BSD10k-v1.2, a +2.12\% improvement over the baseline."
}

@techreport{Lin_JKU_2026,
    Author = "Lin, Pao",
    title = "Heterogeneous Audio Classification with Frozen Audio-Language Embeddings",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "We describe our submission to DCASE 2026 Task 1, Heterogeneous Audio Classification, where each sound is assigned to one of 23 second-level Broad Sound Taxonomy classes and systems are ranked by macro hierarchical-F1 (hF). Like the official baseline, we keep the CLAP audio-text encoder frozen and train small heads on its embeddings. Under matched 5-fold cross-validation on BSD10k-v1.2, our best single model, a gated multimodal head, reaches 78.8 +/- 1.1\% hF, matching the published multimodal baseline. Several additions add little once evaluated carefully. Agreement pseudo-labels mined from BSD35k-CS give no gain that clears seed noise, so we treat them as a neutral data addition. A four-member ensemble, with weights tuned on our validation split, reaches 79.4 +/- 0.7\% hF over three seeds. Of everything we tried, only adding the free-text description to the metadata helped cleanly (+2.2 hF). Finally, the residual error appears strongly tied to label ambiguity: the classes our system confuses most are the ones annotators were least confident labeling, and most of the lost credit lies in cross-family confusions such as crowd speech vs. urban soundscape. We view this label-ambiguity analysis as the main finding of our submission. We submit four systems with different complexity and robustness profiles."
}

@techreport{Liu_CQUPT_2026,
    Author = "Jing, Xiangyu and Luo, Yuandong and Huang, Chaoyong and Liu, Hongqing",
    title = "HAF-CLAP: AHIERARCHICAL-AWAREMULTIMODALCLAPSYSTEMFOR HETEROGENEOUSAUDIO CLASSIFICATION",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report describes the system submitted by Chongqing University of Posts and Telecommunications– Audio Lab (CQUPT AUL) for DCASE 2026 Task 1: Heterogeneous Audio Classifi cation. The proposed system, termed HAF-CLAP, is built upon LAION-CLAP and exploits both acoustic information and textual metadata under the Broad Sound Taxonomy (BST). To adapt pre trained audio-language representations to the target classification task, HAF-CLAP uses a hierarchical-aware multimodal classifica tion framework with audio-text fusion. Several training and infer ence strategies are further applied to improve robustness and pre diction stability. The final submission combines multiple comple mentary HAF-CLAP models. Experimental results on our internal validation split show that the submitted system achieves competitive performance under the hierarchical evaluation metric."
}

@techreport{Primus_CPJKU_2026,
    Author = "Primus, Paul and Widmer, Gerhard",
    title = "CP-JKU Submission to Task 1 of the DCASE 2026 Challenge: LLM Prediction Fusion and Pseudo-Label Training for Heterogeneous Audio Classification",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This report describes our submission to DCASE 2026 Task 1, Heterogeneous Audio Classification. The task requires classifying Freesound audio clips into 23 second-level Broad Sound Taxonomy classes using audio and optional textual metadata. Our system combines pretrained audio-only and audio-text backbones, including PaSST, BEATs, M2D, CP-CLAP, and LAION-CLAP. We use GPT-5.4-mini in two ways: first, to generate metadata summaries for the CLAP text encoders, and second, to estimate class-probability priors from title, tags, and description. These priors are converted into learned class-embedding mixtures and late fused with the backbone representation. To exploit the noisy BSD35k-CS dataset, we use a three-stage training procedure: clean-data training on BSD10k, pseudo-label-based pretraining on BSD35k-CS, and final fine-tuning on BSD10k. The submitted systems are ensembles of selected Stage-3 models and achieve up to 0.834 hierarchical F-score on our BSD10k test split."
}

@techreport{Sharma_IR_2026,
    Author = "Sharma, Nipun and Sharma, Aditya",
    title = "Feature-Centric Late-Fusion Approach for Heterogenous Audio Classification",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This technical report describes the submitted system for DCASE 2026 Task 1, Heterogeneous Audio Classification, which is defined over the Broad Sound Taxonomy with 5 top-level and 23 second-level categories. The proposed approach uses a late-fusion architecture built on frozen foundation-model embeddings, combining CLAP audio embeddings, PANNs audio embeddings, and CLAP text embeddings derived from metadata. Each modality is projected into a shared hidden space and fused with a lightweight Transformer encoder. The resulting sequence is then aggregated via mean-pooling to create a unified representation. Training uses a hierarchical objective that combines fine-level cross-entropy with an auxiliary coarselevel loss, together with label smoothing and class weighting. Evaluation is performed using the hierarchical Fscore adopted for the task, with additional reporting of toplevel and second-level accuracy. For the final submission, logits from seven selected checkpoints were averaged in a checkpoint ensemble to improve robustness and reduce foldspecific variance. The final system utilizes this streamlined three modalities configuration."
}

@techreport{Zhang_XJTLU_2026,
    Author = "Zhang, Peihong and Li, Shengchen",
    title = "Noisy-Label-Aware Multimodal Ensembling with Inference-Safe Candidate Reranking for Heterogeneous Audio Classification",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "We present a noisy-label-aware multimodal system for DCASE 2026 Challenge Task 1, Heterogeneous Audio Classification. The task requires second-level Broad Sound Taxonomy prediction from audio and metadata and is evaluated using macro-averaged hierarchical F-score. Our system addresses the central tension between exploiting large but noisy crowd-sourced supervision and preserving inference safety on the unlabeled evaluation set. It combines LAION-CLAP audio/text embeddings, teacher-weighted soft and hard supervision from BSD35k-CS, a fresh long-run ensemble of 31 checkpointed model families, calibrated probability fusion, and an inference-safe candidate reranker with conservative parent-aware override. All submitted systems are generated from deployable checkpoints rather than replayed development-set artifacts. On a label-blind development-as-evaluation protocol, the primary system obtains 83.5214 hF, improving over our local multimodal baseline by 4.57 absolute hF points."
}

@techreport{Zheng_SCUT_2026,
    Author = "Zheng, Han and Li, Yanxiong",
    title = "MULTIMODAL HATR CLASSIFICATION WITH FROZEN CLAP EMBEDDINGS FOR THE BROAD SOUND TAXONOMY",
    institution = "DCASE2026 Challenge",
    year = "2026",
    month = "June",
    abstract = "This report presents a multimodal audio classifier for DCASE 2026 Task 1, which assigns Freesound recordings to 23 second - level categories of the Broad Sound Taxonomy (BST). The system combines frozen LAION - CLAP embeddings of audio and textual metadata (title, tags, description) with a Hierarchical Audio Tagging and Retrieval (HATR) feed - forward classifier using attention - based fusion. Development performance is measured under five - fold cross - validation on BSD10k - v1.2 and BSD35k - CS (lambda = 0.75). On BSD10k - v1.2, the multimodal model attains macro - averaged hierarchical F - scores of 78.65\% pm 0.52 (hf), compared with 75.88\% pm 0.47 for the audio - only configuration; on BSD35k - CS, the corresponding values are 79.77\% pm 0.60 and 70.19\% pm 0.93, with leaf accuracy reaching 85.32\% pm 0.51. For the challenge evaluation set, five fold models trained on BSD10k - v1.2 are ensembled to produce predictions for all 3246 released test recordings."
}