@techreport{BaiJLESS2023,
    Author = "Huang, Siwei and Bai, Jisheng and Jia, Yafei and Chen, Jianfeng",
    title = "JLESS Submission to DCASE2023 Task7: Foley Sound Synthesis Using Non-Autoagressive Generative Model",
    institution = "School of Marine Science and Technology, Northwestern Polytechnical University, Xi'an, China, LianFeng Acoustic Technologies Co., Ltd. Xi'an, China",
    year = "2023",
    month = "June",
    abstract = "This technical report describes our proposed system for DCASE2023 task7: Foley Sound Synthesis. In our approach, we propose a GAN-based mel-spectrogram synthesis system. we take a Conditional Variational auto-encoder (CVAE) as the generator, which consists of densely-connected dilated convolution blocks, and a simple CNN as the discriminator. The decoder of CVAE synthesizes fake mel-spectrogram resampling from prior noise and class, and the discriminator determines whether it is real or not. Furthermore, we also train a classifier to help CVAE keep class-wise distribution. Finally, the audio is wrapped using the HiFiGAN vocoder."
}

@techreport{ChangHYU2023,
    Author = "Choi, Won-Gook and Chang, Joon-Hyuk",
    title = "HYU Submission For The DCASE 2023 Task 7: Diffusion Probabilistic Model With Adversarial Training For Foley Sound Synthesis",
    institution = "Department of Electronic Engineering, Hanyang University, Seoul, Republic of Korea, Department of Electronic Engineering, Hanyang University, Seoul, Republic of Korea",
    year = "2023",
    month = "June",
    abstract = "This paper is a technical report of the Hanyang University team submission for the DCASE 2023 challenge task 7, Foley Sound Synthesis. The goal of the task is to build a generative model that can synthesize high-quality and various foley sounds: the sounds of dog barking, footsteps, gunshots, keyboards, moving motor vehicles, rainy scenes, and sneezing. The core strategy of the submissions is a diffusion probabilistic model-based acoustic model. Also, we adopted adversarial training on the evidence lower bound (ELBO) of the diffusion model for the higher quality. The submissions did not use any external dataset and achieved lower Frechet audio distance (FAD) scores than the DCASE baseline, except for the sounds of moving motor vehicles."
}

@techreport{ChonGLI2023,
    Author = "Kang, Minsung and Oh, Sangshin and Moon, Hyeongi and Lee, Kyungyun and Chon, Ben Sangbae",
    title = "FALL-E: Gaudio Foley Synthesis System",
    institution = "Gaudio Lab, Inc., Seoul, South Kore",
    year = "2023",
    month = "June",
    abstract = "This paper introduces FALL-E, Gaudio's Foley Synthesis System, which is submitted to the DCASE 2023 Task 7 Foley Synthesis Challenge (Track A). The system employs a cascaded approach comprising low-resolution spectrogram generation, spectrogram super-resolution, and a vocoder. We trained every sound-related model from scratch using our extensive datasets, and we utilized a pre-trained language model. We conditioned the model with dataset-specific texts, enabling it to learn sound quality and recording environment based on the text input. Moreover, we leveraged external language models to improve text descriptions of our datasets and performed prompt engineering for quality, coherence, and diversity. We report the objective measure with respect to the official evaluation set, although our focus is on developing generally working sound generation models beyond the challenge."
}

@techreport{ChunChosun2023,
    Author = "Bang, Chae-Woon and Kim, Nam Kyun and Chun, Chanjun",
    title = "High-Quality Foley Sound Synthesis Using Monte Carlo Dropout",
    institution = "Chosun University, Gwangju, South Korea, Korea Automotive Technology Institute, Gwanjgu, South Korea",
    year = "2023",
    month = "June",
    abstract = "This technical report describes the foley sound synthesis system for DCASE2022 Task7. Here, it aims to creates foley sound, which is widely utilized as various sound effects in multimedia contents. To accomplish this, it uses sound synthesis technique, generating a 4-second audio clip of one of seven classes. Specifically, we fine-tuned the baseline model such that improves the performance. After that, we ensemble the models using Monte Carlo Dropout. The performance of the proposed system was compared with the baseline using Frechet Audio Distance (FAD), which is referred as an audio evaluation metric. As a result, it was confirmed that both the single model and the ensemble model outperform the existing baseline system."
}

@techreport{ChungKAIST2023,
    Author = "Chung, Yoonjin and Lee, Junwon and Nam, Juhan",
    title = "Foley Sound Synthesis In Waveform Domain With Diffusion Model",
    institution = "Graduate School of AI, KAIST, Graduate School of Culture Technology, KAIST",
    year = "2023",
    month = "June",
    abstract = "Foley sound synthesis becomes an important task due to the growing popularity of multi-media content, which is an industrial usecase of general audio synthesis. We propose a diffusion-based model that generates class-conditioned general audio in a classifier-free guidance manner as a participant of DCASE 2023 challenge task 7[1]. Our model follows a UNet-like structure while incorporating LSTM[2] inside the encoder block. We demonstrate the FAD (Frechet Audio Distance) scores of generated results for each 7 sound class respectively."
}

@techreport{GuanHEU2023,
    Author = "Fan, Shitong and Zhu, Qiaoxi and Xiao, Feiyang and Lan, Haiyan and Wang, Wenwu and Guan1, Jian",
    title = "Foley Sound Synthesis With AudioLDM For DCASE2023 Task 7",
    institution = "Group of Intelligent Signal Processing (GISP), College of Computer Science and Technology, Harbin Engineering University, Harbin, China, Centre for Audio, Acoustic and Vibration (CAAV), University of Technology Sydney, Ultimo, Australia, Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, Guildford, UK",
    year = "2023",
    month = "June",
    abstract = "This report describes our submission for DCASE2023 Challenge Task 7, a system for foley sound synthesis. Our system is based on AudioLDM, which offers high generation quality and computational efficiency for the text-to-audio task. Experiments are conducted on the dataset of DCASE2023 Challenge Task 7. The Fr\'{e}chet audio distance (FAD) between the sound generated by our system and the actual sound sample is 5.120 in the category “DogBark” and 8.102 in the category “Rain”, better than the baseline with an FAD 7.256 and an FAD 4.901 distance closer to the actual samples, respectively."
}

@techreport{JungKT2023,
    Author = "Chung, Hae Chun and Lee, Yuna and Jung, Jae Hoon",
    title = "Foley Sound Synthesis Based On GAN Using Contrastive Learning Without Label Information",
    institution = "KT Corporation, Republic of Korea",
    year = "2023",
    month = "June",
    abstract = "Sound effects used in radio or movies, such as foley sound, have been difficult to create without the help of experts. Furthermore, in the field of audio synthesis, the field of speech has been actively progressed, but there has been no research on audio sounds that can be obtained in real life. In this technical report, We present our submission system for DCASE2023 Task7: Foley-sound synthesis. We participate in track B, which forbids the usage of external resources. We propose a framework that employ the loss function of ContraGAN and C-SupConGAN based on structure of Self-Attention GAN (SAGAN). Our final system achieves outperforming the baseline performance by a large margin."
}

@techreport{KamathNUS2023,
    Author = "Kamath, Purnima and Islam, Tasnim Nishat and Gupta, Chitralekha and Wyse, Lonce and Nanayakkara, Suranga",
    title = "DCASE Task-7: StyleGAN2-Based Foley Sound Synthesis",
    institution = "National University of Singapore, Singapore and Bangladesh University of Engineering and Technology, Bangladesh and Universitat Pompeu Fabra, Barcelona, Spain",
    year = "2023",
    month = "June",
    abstract = "For the DCASE 2023 Task 7 (Track B), Foley Sound Synthesis, we submit two systems, (1) a StyleGAN conditioned on the class ID, and (2) an ensemble of StyleGANs each trained unconditionally on each class separately. We quantitatively find that both systems out-perform the task 7 baseline models in terms of FAD Scores. Given the high inter-class and intra-class variance in the development datasets, the system conditioned on class ID is able to generate a smooth and a homogeneous latent space indicated by the subjective quality of its generated samples. The unconditionally trained ensemble generates more categorically recognizable samples than system 1, but tends to generate more instances of out-of-distribution or noisy samples."
}

@techreport{KeunwooGLI2023,
    Author = "Choi, Keunwoo and Im, Jaekwon and Heller, Laurie and McFee, Brian and Imoto, Keisuke and Okamoto, Yuki and Lagrange, Mathieu and Takamichi, Shinnosuke",
    title = "Foley Sound Synthesis at the DCASE 2023 Challenge",
    institution = "Gaudio Lab, Inc., Seoul, South Korea, KAIST and Daejeon, South Korea and Carnegie Mellon University, Pennsylvania, USA and New York University, New York, USA, and Doshisha University, Kyoto, Japan and Ritsumeikan University, Kyoto, Japan and CNRS, Ecole Centrale Nantes, Nantes Universite, Nantes, France and The University of Tokyo, Tokyo, Japan",
    year = "2023",
    month = "June",
    abstract = "The addition of Foley sound effects during post-production is a common technique used to enhance the perceived acoustic properties of multimedia content. Traditionally, Foley sound has been produced by human Foley artists, which involves manual recording and mixing of sound. However, recent advances in sound synthesis and generative models have generated interest in machine-assisted or automatic Foley synthesis techniques. To promote further research in this area, we have organized a challenge in DCASE 2023: Task 7 - Foley Sound Synthesis. Our challenge aims to provide a standardized evaluation framework that is both rigorous and efficient, allowing for the evaluation of different Foley synthesis systems. Through this challenge, we hope to encourage active participation from the research community and advance the state-of-the-art in automatic Foley synthesis. In this technical report, we provide a detailed overview of the Foley sound synthesis challenge, including task definition, dataset, baseline, evaluation scheme and criteria, and discussion."
}

@techreport{LeeMARG2023,
    Author = "Kim, Kyungsu and Lee, Jinwoo and Kim, Hayoon and Lee, Kyogu",
    title = "Conditional Foley Sound Synthesis With Limited Data: Two-Stage Data Augmentation Approach With StyleGAN2-ADA",
    institution = "Seoul National University Department of Intelligence and Information",
    year = "2023",
    month = "June",
    abstract = "This report introduces an audio synthesis system designed to tackle the task of Foley Sound Synthesis in the Detection and Classification of Acoustic Scenes and Events (DCASE) 2023 challenge. Our proposed system comprises an ensemble of a baseline model and StyleGAN2-ADA. To optimize the system with the limited data without relying on external datasets and pretrained systems, we propose a two-stage data augmentation strategy. This approach involves augmenting input waveforms to expand the size of the training dataset, as well as employing adaptive discriminator augmentation (ADA) to alleviate overfitting of discriminator and ensure stable training. Experimental results demonstrate that our proposed ensemble system achieves an FAD (Fr\'{e}chet Audio Distance) of 5.84 on the evaluation dataset."
}

@techreport{Leemaum2023,
    Author = "Lee1, Junhyeok and Nam, Hyeonuk and Park, Yong-Hwa",
    title = "VIFS: An End-To-End Variational Inference For Foley Sound Synthesis",
    institution = "maum.ai Inc., Republic of Korea and Korea Advanced Institute of Science and Technology, Republic of Korea",
    year = "2023",
    month = "June",
    abstract = "Foley sound synthesis (FSS) is a task to generate a sound for specific conditions. In this work, FSS is defined as a ”category-to-sound” problem, which is generating various sounds for a given category. To solve this diversity problem, we adopt VITS, a text-to-speech (TTS) model with variational inference. In addition, we apply various techniques from speech synthesis including PhaseAug and Avocodo. Different from TTS models which generate short pronunciation from phonemes and a speaker identity, the category-to-sound problem requires to generate diverse sounds just from a category class. To compensate this difference between TTS and category-to- sound while maintaining consistency within each inference, we heavily modified the prior encoder to enhance consistency with posterior latent variables. This introduced additional gaussian on prior encoder which promotes variance within the category. With these modifications, we propose VIFS, variational inference for end-toend Foley sound synthesis, which is able to generate high-quality sounds with diversity."
}

@techreport{PillayCMU2023,
    Author = "Pillay, Ashwin and Betko, Sage and Liloia, Ari and Chen, Hao and Shah, Ankit",
    title = "DCASE Task 7: Foley Sound Synthesis",
    institution = "Carnegie Mellon University, Pittsburgh, USA",
    year = "2023",
    month = "June",
    abstract = "Foley sound synthesis refers to the creation of realistic, diegetic sound effects for a piece of media, such as film or radio. We propose building a deep learning system for Task 7 of the DCASE 2023 challenge that can generate original mono audio clips belonging to one of seven foley sound categories. Our training dataset consists of 4,850 sound clips from the UrbanSound8K, FSD50K, and BBC Sound Effects datasets. We aim to better the subjective and objective quality of generated sounds by passing as much meaningful information about the input data into latent representations as possible. The primary innovation in our submission is the change from using melspectrograms to using CEmbeddings (combined embeddings), which are input to the VQ-VAE and consist of melspectrograms concatenated with latent representations of audio produced by a pre-trained MERT model. Our submission to track A utilizes a pre-trained MERT model; as such, PixelSNAIL was trained on CEmbeddings. Our submission to track B utilizes PixelSNAIL retrained only on melspectrograms. Our code can be found here: https://github.com/ankitshah009/ foley-sound-synthesis\_DCASE\_2023."
}

@techreport{QianbinBIT2023,
    Author = "Qi, Anbin",
    title = "Auto-Bit for DCASE2023 Task7 Technical Reports: Assemble System of BitDiffusion and PixelSNAIL",
    institution = "School Information and Electronics, Beijing Institute of Technology, Beijing, China",
    year = "2023",
    month = "June",
    abstract = "This paper is a technical report on DCASE TASK7, which proposes using different methods and models for sound synthesis tasks in different scene events. For dogbark and sneeze cough, a non autoregressive model based on conditional generation bit-diffusion was used for sound synthesis. For the other five types of sounds, a autoregressive model based PixelSnail was used."
}

@techreport{QianXuBIT2023,
    Author = "Zhang, Haojie and Qian, Kun and Shen, Lin and Li, Lujundong and Xu, Kele and Hu, Bin",
    title = "From Noise To Sound: Audio Synthesis Via Diffusion Models",
    institution = "Key Laboratory of Brain Health Intelligent Evaluation and Intervention, Ministry of Education (Beijing Institute of Technology), P. R. China, School of Medical Technology, Beijing Institute of Technology, P. R. China, National University of Defense Technology, P. R. China",
    year = "2023",
    month = "June",
    abstract = "In this technical report, we describe our submission system for DCASE2023 Task 7: Foley Sound Synthesis (Track B). A Sound Pixelate Diffuse model is proposed to realize foley sound synthesis. The model includes data format conversion and synthesising audio through the diffusion model. The Synthesised audio are evaluated on DCASE2023 Task 7 Eval FAD evaluation set and the best FAD score of all categories is 8.429."
}

@techreport{ScheiblerLINE2023,
    Author = "Scheibler, Robin and Hasumi, Takuya and Fujita, Yusuke and Komatsu, Tatsuya and Yamamoto, Ryuichi and Tachibana, Kentaro",
    title = "Class-Conditioned Latent Diffusion Model For DCASE 2023 Foley Sound Synthesis Challenge",
    institution = "LINE Corporation, Tokyo, Japan",
    year = "2023",
    month = "June",
    abstract = "This report describes our submission to DCASE 2023 Task7: Foley sound synthesis challenge. We use a latent diffusion model (LDM) that generates a latent representation of audio conditioned on a specified audio class, a variational autoencoder that converts the latent representation to a mel-spectrogram, and a universal neural vocoder based on HIFI-GAN that reconstructs a natural waveform from the mel-spectrogram. We trained the LDM using development set with its audio class indices as conditioners for generating class-specific latent representations."
}

@techreport{WendnerJKU2023,
    Author = "Wendner, Timo and Hu, Patricia and Jadidi, Tara and Neuhauser, Alexander",
    title = "Audio Diffusion For Foley Sound Synthesis",
    institution = "Johannes Kepler University, Linz, Austria",
    year = "2023",
    month = "June",
    abstract = "This technical report describes our approach for Task 7 (Foley Sound Synthesis), Track B (using no external resources other than the ones provided) of the DCASE2023 Challenge. This work was carried out as part of an elective course in the Artificial Intelligence curriculum at Johannes Kepler University Linz by a student group. We use an ensemble of U-Net based diffusion models for waveform generation in seven predefined sound categories. We apply gain reduction to normalize and time shifting to augment the provided training data and test different noise schedulers and U-Net architectures. Applying different training strategies, we achieve competitive results for the majority of the sound classes while being more parameter efficient and allowing end-to-end generation on audio waveforms. Evaluated on the task's evaluation metric, i.e., the mean FAD score over all classes, we achieve a final score of 12.42 as compared to the score of the challenge baseline model of 9.68."
}

@techreport{XieSJTU2023,
    Author = "Xie, Zeyu and Xu, Xuenan and Li, Baihan and Wu, Mengyue and Yu, Kai",
    title = "The X-LANCE System For DCASE2023 Challenge Task 7: Foley Sound Synthesis Track B",
    institution = "MoE Key Lab of Artificial Intelligence X-LANCE Lab, Department of Computer Science and Engineering AI Institute, Shanghai Jiao Tong University, Shanghai, China",
    year = "2023",
    month = "June",
    abstract = "This report describes the system submitted to the Detection and Classification of Acoustic Scenes and Events (DCASE) 2023 challenge Task 7: foley sound synthesis track B. We first train a VQVAE model to learn the discrete representation of the audio spectrogram. Then an auto-regressive model is trained to predict discrete tokens based on input conditions. Finally, a trained vocoder converts the generated spectrogram into a waveform, where the spectrogram is restored from predicted tokens by VQ-VAE decoder. To achieve higher accuracy, fidelity and diversity, we introduce some training schemes, including (1) a discriminator model to filter audio; (2) mixup method for data augmentation; (3) clustering methods for better training. Our best system achieved a FAD score of 6.99 averaging on all categories."
}

@techreport{YiSURREY2023,
    Author = "Yuan, Yi and Liu, Haohe and Liu, Xubo and Kang, Xiyuan and Plumbley, Mark D. and Wang, Wenwu",
    title = "Latent Diffusion Model Based Foley Sound Generation System For DCASE Challenge 2023 Task 7",
    institution = "University of Surrey, Guildford, United Kingdom",
    year = "2023",
    month = "June",
    abstract = "Foley sound generation aims to synthesise the background sound for multimedia content, which involves computationally modelling sound effects with specialized techniques. In this work, we proposed a diffusion-based generative model for DCASE 2023 challenge task 7: Foley Sound Synthesis. The proposed system is based on AudioLDM, which is a diffusion-based text-to-audio generation model. To alleviate the data scarcity of task 7 training set, our model is initially trained with large-scale datasets and downstream into this DCASE task via transfer learning. We have observed that the feature extracted by the encoder can significantly affect the performance of the generation model. Hence, we improve the results by leveraging the input label with related text embedding features obtained by a large language model, i.e.,contrastive language-audio pretraining (CLAP). In addition, we utilize a filtering strategy to further refine the output, i.e. by selecting the best results from the candidate clips generated in terms of the similarity score between the sound and target labels. The overall system achieves a Fr\'{e}chet audio distance (FAD) score of 4.765 on average among all seven different classes, outperforming the baseline system which achieves a FAD score of 9.7."
}