@techreport{chen2021_t6,
    Author = "Chen, Zhiwen and Zhang, Dawei and Wang, Jun and Deng, Feng",
    title = "Audio Captioning With Meshed-Memory Transformer",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "Automated audio captioning is the task of describing the audio content of a given audio signal in natural language. Through advancing for years, transformer-based language models are widely used in audio captioning. However, most architectures based on transformer, cannot learn prior knowledge between samples well, leading to worse text decoding. For better acoustic event and language modeling, a sequence-to-sequence model is proposed which consists of a CNN-based encoder, a memory-augmented refiner and a meshed decoder. The proposed architecture refines a multi-level representation of the relationships between audio features integrating learned a priori knowledge. At decoding stage, it exploit low- and high-level features with a mesh-like connectivity. Experiments show that the proposed model can achieve a SPIDEr score of 0.2645 on Clotho V2 dataset."
}

@techreport{eren2021_t6,
    Author = {\"{O}zkaya Eren, Ay\c{s}eg\"{u}l and Sert, Mustafa},
    title = "Audio Captioning Using Sound Event Detection",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This technical report proposes an audio captioning system for DCASE 2021 Task 6 audio captioning challenge. Our proposed model is based on an encoder-decoder architecture with bi-directional Gated Recurrent Units (BiGRU) using pretrained audio features and sound event detection. A pretrained neural network (PANN) is used to extract audio features and Word2Vec is selected with the aim of extracting word embeddings from the audio captions. To create semantically meaningful captions, we extract sound events from the audio clips and feed the encoder-decoder architecture with sound events in addition to PANNs features. Our experiments on the Clotho dataset show that our proposed method significantly achieves better results than the challenge baseline model across all evaluation metrics."
}

@techreport{gebhard2021_t6,
    Author = {Gebhard, Alexander and Triantafyllopoulos, Andreas and Baird, Alice and Schuller, Bj\"{o}rn},
    title = "An Automated Audio Captioning Approach Utilising a ResNet-based Encoder",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "In this report, we present our submission system to TASK6 of the DCASE2021 Challenge. The main module is based on the baseline architecture for the automated audio captioning (AAC) task, which was provided by the challenge organisers. We exchange the encoder part of the baseline architecture and replace it by a Residual Neural Network (ResNet)-18 encoder adapted to the AAC task. Results from our proposed architecture have shown an average increase of 35.7\\% over the baseline system, reaching a BLEU1 score of 0.449 on the development set, demonstrating the effectiveness of the proposed encoder for this task."
}

@techreport{labbe2021_t6,
    Author = "Labb\'{e}, Etienne and Pellegrini, Thomas",
    title = "{IRIT-UPS} {DCASE} 2021 Audio Captioning System",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This document provides a description of our seq-to-seq models used for audio captioning in the task 6 of the DCASE 2021 challenge. Four submissions were made with two different models, a “Listen- Attend-Tell” and a “CNN-Tell”, and two different algorithms for inference, greedy and beam search."
}

@techreport{liu2021_t6,
    Author = "Yang, Liu and Sijun, Bi",
    title = "The {DCASE2021} Challenge Task 6 System : Automated Audio Caption",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This technical report describes the system participating to the Automated Audio Captioning (DCASE) 2021 Challenge, Task 6: automated audio captioning. In this work, We employ several learnable stack CNNs to extract audio features in the encoder layer, meanwhile, we employ the decoder of the widely used Transformer structure to generate captions. For optimize system, we use the sentence-level cosine loss function and crossentropy loss. The experimental results show that our system could achieve the SPIDEr of 0.166 on the evaluation split of the Clotho dataset."
}

@techreport{narisetty2021_t6,
    Author = "Narisetty, Chaitanya and Hayashi, Tomoki and Ishizaki, Ryunosuke and Watanabe, Shinji and Takeda, Kazuya",
    title = "Leveraging State-of-the-art {ASR} Techniques to Audio Captioning",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This report presents a summary of our submission to the 2021 DCASE challenge Task 6: Automated Audio Captioning. Our ap- proach to this task is derived from state-of-the-art ASR techniques available in the ESPNet toolkit. Specifically, we train a convolution-augmented Transformer (Conformer) model to generate captions from input acoustic features in an end-to-end manner. In addition to the prescribed challenge dataset: Clotho-v2, we also augment the AudioCaps external dataset. To overcome the limited availability of training data, we further incorporate the Audioset-tags and audio-embeddings obtained by pretrained audio neural networks (PANNs) as an auxiliary input to our model. An ensemble of models trained over various architectures and input embeddings is selected as our final submission system. Experimental results indicate that our models achieve a SPIDEr score of 0.224 and 0.246 on the development-validation and development-evaluation sets respectively."
}

@techreport{won2021_t6,
    Author = "Won, Hyejin and Kim, Baekseung and Kwak, Il-Youp and Lim, Changwon",
    title = "{CAU} Submission to {DCASE} 2021 Task6: Transformer Followed by Transfer Learning for Audio Captioning",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This report proposes an automated audio captioning model for the 2021 DCASE audio captioning challenge. In this challenge, a model is required to generate natural language descriptions of a given audio signal. We use pre-trained models trained using AudioSet data, a large-scale dataset of manually annotated audio events. The large amount of audio events data would help capturing important au- dio feature representation. To make use of the learned feature from AudioSet data, we explored several transfer learning approaches. Our proposed sequence-to-sequence model consists of a CNN14 or ResNet54 encoder and a Transformer decoder. Experiments show that the proposed model can achieve a SPIDEr score of 0.246 and 0.285 on audio captioning performance."
}

@techreport{xiao2021_t6,
    Author = "Xiao, Feiyang and Guan, Jian and Kong, Qiuqiang",
    title = "Automated Audio Captioning With {MLP}-mixer and Pre-trained Encoder",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This technical report describes the submission from the Group of Intelligent Signal Processing (GISP) for Task6 of DCASE2021 challenge (automated audio captioning). Our audio captioning system is based on the sequence-to-sequence autoencoder model. Previous recurrent neural network (RNN) and Transformer based methods just perceive the time dimension information but ignore the frequency information. To utilize both time dimension and frequency dimension information, multi-layer perceptrons mixer (MLP-Mixer) is used as the encoder. For caption prediction, a Transformer decoder structure is used as the decoder. No extra data is employed. In addition, to highlight the content information, we use a pre-trained encoder with multi-label content information. The experimental results show that our system can achieve the SPIDEr of 0.144 (official baseline: 0.051) on the evaluation split of the Clotho dataset. In addition, comparing with Transformer methods, our system has fewer training time."
}

@techreport{xinhao2021_t6,
    Author = "Mei, Xinhao and Huang, Qiushi and Liu, Xubo and Chen, Gengyun and Wu, Jingqian and Wu, Yusong and Zhao, Jinzheng and Li, Shengchen and Ko, Tom and Tang, H. Lilian and Shao, Xi and Plumbley, Mark D. and Wang, Wenwu",
    title = "An Encoder-Decoder Based Audio Captioning System With Transfer and Reinforcement Learning for {DCASE} Challenge 2021 Task 6",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "Audio captioning aims to use natural language to describe the content of audio data. This technical report presents an automated audio captioning system submitted to Task 6 of the DCASE 2021 challenge. The proposed system is based on an encoder-decoder architecture, consisting of a convolutional neural network (CNN) encoder and a Transformer decoder. We further improve the system with two techniques, namely, pre-training the model via transfer learning techniques, either on upstream audio-related tasks or large in-domain datasets, and incorporating evaluation metrics into the optimization of the model with reinforcement learning techniques, which help address the problem caused by the mismatch between the evaluation metrics and the loss function. The results show that both techniques can further improve the performance of the captioning system. The overall system achieves a SPIDEr score of 0.277 on the Clotho evaluation set, which outperforms the top-ranked system from the DCASE 2020 challenge."
}

@techreport{xu2021_t6,
    Author = "Xu, Xuenan and Xie, Zeyu and Wu, Mengyue and Yu, Kai",
    title = "The {SJTU} System for {DCASE2021} Challenge Task 6: Audio Captioning Based on Encoder Pre-training and Reinforcement Learning",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This report proposes an audio captioning system for the Detection and Classification of Acoustic Scenes and Events (DCASE) 2021 challenge task Task 6. Our audio captioning system consists of a 10-layer convolution neural network (CNN) encoder and a temporal attentional single layer gated recurrent unit (GRU) decoder. In this challenge, there is no restriction on the usage of external data and pre-trained models. To better model the concepts in an audio clip, we pre-train the CNN encoder with audio tagging on AudioSet. After standard cross entropy based training, we further fine-tune the model with reinforcement learning to directly optimize the evaluation metric. Experiments show that our proposed system achieves a SPIDEr of 28.6 on the public evaluation split without ensemble."
}

@techreport{ye2021_t6,
    Author = "Ye, Zhongjie and Wang, Helin and Yang, Dongchao and Zou, Yuexian",
    title = "Improving the Performance of Automated Audio Captioning via Integrating the Acoustic and Textual Information",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This technical report describes an automated audio captioning (AAC) model for the the Detection and Classification of Acoustic Scenes and Events (DCASE) 2021 Task 6 Challenge. In order to utilize more acoustic and textual information, we propose a novel sequence-to-sequence model named KPE-MAD, with a keyword pre-trained encoder and a multi-modal attention decoder. For the encoder, we use pre-trained classification model on the AudioSet dataset, and finetune it with keywords of nouns and verbs as labels. In addition, a multi-modal attention module is proposed to integrate the acoustic and textual information in the decoder. Our single model achieves the SPIDEr score of 0.279 in the evaluation splits. And our best ensemble model by optimizing CIDEr-D via the reinforcement learning, achieves the SPIDEr score of 0.291. Our code1 and models will be released after the competition."
}

@techreport{yuan2021_t6,
    Author = "Yuan, Weiqiang and Han, Qichen and Liu, Dong and Li, Xiang and Yang, Zhen",
    title = "The {DCASE} 2021 Challenge Task 6 System: Automated Audio Captioning With Weakly Supervised Pre-traing and Word Selection Methods",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "July",
    abstract = "This technical report describes the system participating to the De- tection and Classification of Acoustic Scenes and Events (DCASE) 2021 Challenge, Task 6: automated audio captioning. We use encoder-decoder modeling framework for audio understanding and caption generation. Our solution focuses on solving two problems in automated audio captioning: data insufficiency and word selection indeterminacy. As the amount of training data is limited, we collect large-scale weakly labeled dataset from Web with heuristic methods. Then we pre-train the encoder-decoder models with this dataset followed by fine-tuning on Clotho dataset. To solve the word selection indeterminacy problem, we use keywords extracted from captions of similar audios and audio event tags produced by pre-trained audio tagging models to guide words generation in decoding stage. We tested our submissions using the development-testing dataset. Our best submission achieved 31.8 SPIDEr score."
}