@techreport{Chung2024_t9,
    Author = "Chung, Hae Chun and Jung, Jae Hoon",
    title = "LANGUAGE-QUERIED AUDIO SOURCE SEPARATION ENHANCED BY EXPANDED LANGUAGE-AUDIO CONTRASTIVE LOSS",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "This technical report outlines the efforts of KT Corporation's Acoustic Processing Project for addressing language-queried audio source separation (LASS), DCASE 2024 Challenge Task 9. The objective of this work is to separate arbitrary sound sources using a text description of the desired source. We propose three systems, each with the same model architecture but different training methods. These systems use the FLAN-T5 model as the text encoder and the ResUNet model as the separator. To train these systems, we introduced three loss functions: L1 loss in the time domain, multi-scale mel-spectrogram loss in the frequency domain, and contrastive loss, with a loss balancer to stabilize the training. Utilizing the Contrastive Language-Audio Pre-training (CLAP) model, we designed three contrastive losses: audio-to-text (A2T-CL), audio-to-audio (A2A-CL), and audio-to-multi (A2M-CL). The first system was trained with A2T-CL, the second with both A2A-CL and A2T-CL, and the third with A2M-CL. These systems achieved signal-to-distortion ratio (SDR) of 7.030, 7.124, and 7.136, respectively, showing nearly a 30\\% improvement over the baseline SDR of 5.708 provided by the challenge."
}

@techreport{Lee2024_t9,
    Author = "Lee, Do Hyun and Song, Yoonah and Kim, Hong Kook",
    title = "PERFORMANCE IMPROVEMENT OF LANGUAGE-QUERIED AUDIO SOURCE SEPARATION BASED ON CAPTION AUGMENTATION FROM LARGE LANGUAGE MODELS FOR DCASE CHALLENGE 2024 TASK 9",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "We present a prompt-engineering-based text-augmentation approach applied to a language-queried audio source separation (LASS) task. To enhance the performance of LASS, the proposed approach utilizes large language models (LLMs) to generate multiple captions corresponding to each sentence of the training dataset. To this end, we first perform experiments to identify the most effective prompts for caption augmentation with a smaller number of captions. A LASS model trained with these augmented captions demonstrates improved performance on the DCASE 2024 Task 9 validation set compared to that trained without augmentation. This study highlights the effectiveness of LLM-based caption augmentation in advancing language-queried audio source separation."
}

@techreport{Romaniuk2024_t9,
    Author = "Romaniuk, Michal and Krzywdziak, Justyna",
    title = "SRPOL submission to DCASE 2024 challenge task 9: Modeling real and imaginary components, Mixit and SDR based loss",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "We present our solution to the DCASE 2024 challenge task 9 (Language-Queried Audio Source Separation). Our solution is based on the official baseline, with training dataset including FSD50k, Clotho and additionally extended with AudioCaps. We show that the additional data improve results throughout the training process. We explore changing the ratio masking method from spectrogram amplitude and phase to individual masks for real and imaginary components. We also investigate how different losses, such as Mixit loss and SDR based loss, affect the training process."
}

@techreport{Xiao2024_t9,
    Author = "Xiao, Feiyang and Wang, Wenbo and Xu, Dongli and Qi, Shuhan and Zhang, Kejia and Zhu, Qiaoxi and Guan, Jian",
    title = "LANGUAGE-QUERIED AUDIO SOURCE SEPARATION WITH GPT-BASED TEXT AUGMENTATION AND IDEAL RATIO MASKING",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "This technical report details our submission systems for Task 9 (language-queried audio source separation) of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2024 Challenge. Our four proposed systems utilize the large language model GPT-4 for data augmentation and apply the ideal ratio masking strategy in the latent feature space of the text-to-audio generation model, AudioLDM. Additionally, our systems incorporate the pre-trained language-queried audio source separation model, AudioSep-32K, which leverages extensive pre-training on large-scale data to separate audio sources based on text queries. Experimental results demonstrate that our systems achieve better separation performance compared to the official baseline method on objective metrics. Furthermore, we introduce a novel evaluation metric, audio-text similarity (ATS), which measures the semantic similarity between the separated audio and the text query without requiring a reference target audio signal."
}

@techreport{Yin2024_t9,
    Author = "Yin, Han and Bai, Jisheng and Wang, Mou and Chen, Jianfeng",
    title = "LANGUAGE-QUERIED AUDIO SOURCE SEPARATION VIA RESUNET WITH DPRNN",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "This report presents our submitted systems for the task 9 of DCASE challenge: language-queried audio source separation (LASS). LASS is the task of separating arbitrary sound sources using textual descriptions of the desired source, also known as ''separate what you describe''. Specifically, we first incorporate a dual-path recurrent neural network (DPRNN) block into ResUNet, which is significantly beneficial for improving the separation performance. Then, we trained the proposed model using a large number of public datasets, including Clotho, FSD50K, Audiocaps, Auto-ACD, and Wavcaps. We trained the proposed model at 16 kHz and 32 kHz respectively, and the 32 kHz model achieved the best separation performance with an SDR of 8.191 dB on the validation set, which is 2.483 dB higher than the challenge baseline."
}