Paper deep dive
Environmental Sound Deepfake Detection Using Deep-Learning Framework
Lam Pham, Khoi Vu, Dat Tran, Phat Lam, Vu Nguyen, David Fischinger, Alexander Schindler, Martin Boyer, Son Le
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/26/2026, 11:44:23 PM
Summary
The paper proposes a deep-learning framework for Environmental Sound Deepfake Detection (ESDD), distinguishing between sound scene and sound event detection. The authors demonstrate that fine-tuning pre-trained models, specifically the WavLM/BEATs model, is more effective than training from scratch. A key contribution is a three-stage training strategy involving A-Softmax, Contrastive, and Central losses. Experimental results on the EnvSDD and ESDD-Challenge-TestSet datasets show that the best performing model (BEATs-Finetune+MLP) achieves 0.98 accuracy on the EnvSDD test subset and demonstrates high generalizability in cross-dataset testing.
Entities (8)
Relation Signals (4)
EnvSDD → contains → Sound Scene
confidence 100% · Among the environmental sound, it comprises two main audio types: sound events... and sound scene.
BEATs → usedforesdd → Environmental Sound Deepfake Detection (ESDD)
confidence 100% · we also evaluate the pre-trained BEATs [13] model... to adapt with the downstream ESDD task.
Three-stage training strategy → improves → Environmental Sound Deepfake Detection (ESDD)
confidence 95% · The proposed three-stage training strategy to enhance the task performance of Acoustic Scene Classification and Acoustic Event Detection.
Gammatone (GAM) → outperforms → MEL
confidence 90% · GAM spectrogram outperforms MEL and CQT spectrograms when using the same network architecture EfficientNetB1.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In this paper, we propose a deep-learning framework for environmental sound deepfake detection (ESDD) -- the task of identifying whether the sound scene and sound event in an input audio recording is fake or not. To this end, we conducted extensive experiments to explore how individual spectrograms, a wide range of network architectures and pre-trained models, ensemble of spectrograms or network architectures affect the ESDD task performance. The experimental results on the benchmark datasets of EnvSDD and ESDD-Challenge-TestSet indicate that detecting deepfake audio of sound scene and detecting deepfake audio of sound event should be considered as individual tasks. We also indicate that the approach of finetuning a pre-trained model is more effective compared with training a model from scratch for the ESDD task. Eventually, our best model, which was finetuned from the pre-trained WavLM model with the proposed three-stage training strategy, achieve the Accuracy of 0.98, F1 Score of 0.95, AuC of 0.99 on EnvSDD Test subset and the Accuracy of 0.88, F1 Score of 0.77, and AuC of 0.92 on ESDD-Challenge-TestSet dataset.
Tags
Links
- Source: https://arxiv.org/abs/2604.19652v1
- Canonical: https://arxiv.org/abs/2604.19652v1
Trouble viewing inline? Open PDF directly →
Full Text
25,940 characters extracted from source content.
Expand or collapse full text
Environmental Sound Deepfake Detection Using Deep-Learning Framework Lam Pham*, Khoi Vu*, Dat Tran*, Phat Lam, Vu Nguyen, David Fischinger, Alexander Schindler, Martin Boyer, Son Le† Abstract— In this paper, we propose a deep-learning frame- work for environmental sound deepfake detection (ESDD) – the task of identifying whether the sound scene and sound event in an input audio recording is fake or not. To this end, we conducted extensive experiments to explore how individual spectrograms, a wide range of network architectures and pre- trained models, ensemble of spectrograms or network archi- tectures affect the ESDD task performance. The experimental results on the benchmark datasets of EnvSDD and ESDD- Challenge-TestSet indicate that detecting deepfake audio of sound scene and detecting deepfake audio of sound event should be considered as individual tasks. We also indicate that the approach of finetuning a pre-trained model is more effective compared with training a model from scratch for the ESDD task. Eventually, our best model, which was finetuned from the pre-trained WavLM model with the proposed three-stage training strategy, achieve the Accuracy of 0.98, F1 Score of 0.95, AuC of 0.99 on EnvSDD Test subset and the Accuracy of 0.88, F1 Score of 0.77, and AuC of 0.92 on ESDD-Challenge-TestSet dataset Items— Spectrogram, Deepfake, Audio Embedding, Finetun- ing, Pre-trained Model. I. INTRODUCTION AI-based generation systems are now able to create very realistic audio that are applied for a wide range of applica- tions such as media production, AI-based music composer, audio-based chatbot, etc. However, these generation systems also pose potential risks when generated audio, referred to as deepfake audio, is used for criminal purposes. While the research community on audio mainly focuses on detecting deepfakes in speech or singing voice, deepfake audio of en- vironmental sounds has recently gained the attention. Indeed, the first four audio datasets, which were proposed for the ESDD task, have been recently published: FoleySound [1] (2023), SceneFake [2] (2024), EnvSDD [3](2025), and ESDD-Challenge-TestSet [4]. Among these datasets, only EnvSDD [3] comprises fake audio for both sound scene and sound event. This dataset also presents the largest number of audio samples and recording time. Regarding the environmental sound, it comprises two main audio types: sound events (e.g., the sound of dog bagging, the sound of siren, etc.) and sound scene (e.g., sound on a bus, sound in a park). In comparison to speech, both sound L. Pham, D. Fischinger, A. Schindler, and M. Boyer are with Center for Digital Safety & Security, Austrian Institute of Technology (AIT), Austria D. Tran and K. Vu are with FPT University, Vietnam S. Le is with Ton Duc Thang University, Vietnam P. Lam is with HCM University of Technology, Vietnam (†) Corresponding author. (*) Main and equal contribution into the paper. scenes and sound events present unstructured in form. Sound scenes are closely similar to noise in certain contexts and normally are presented in long-time durations. Meanwhile, there are various sound events in real life which spread in a wide range of frequency bands and are presented in sort-time durations. As environmental sounds present different acous- tic characteristics compared with human speech, proposed methods for detecting human speech and singing deepfakes may be less effective and need to be evaluated. However, a few of papers [5], [6] proposed for the ESDD task have been published and only focus on deepfake audio of sound event. As the ESDD task has been drawn attention and the audio research community presents a lack of evaluation on sound scene and sound event independently, these concerns inspire us to evaluate the role of different spectrograms and various deep neural network architectures with respect to the ESDD task for both sound event & sound scene. Given the experimental results, we indicate whether one deep-learning model is possible to well performs the ESDD task on both sound event and sound scene, or each type of audio requires an individual model. We evaluate our proposed models in this paper on the benchmark dataset of EnvSDD [3] which presents both deepfake audio of sound event and sound scene. We also present a cross-dataset test with the ESDD-Challenge-TestSet [4] to evaluate the the generality of proposed models. I. PROPOSED DEEP-LEARNING FRAMEWORK A. The High-Level Architecture of The proposed Deep- learning Framework We first propose a deep-learning framework for envi- ronmental sound deepfake detection (ESDD). The high- level architecture of our proposed framework is described in Fig. 1. In particular, the audio is first transformed into spectrograms. To deal with the unbalanced issue between fake and real audio recordings, Mixup [7], [8], is applied as the online data augmentation. Then, original and Mixup spectrograms are fed into a deep neural network architecture, referred to as the backbone, to extract the audio embedding (i.e., an audio embedding is a vector-based presentation). Finally, the audio embedding is explored by a multilayer perception (MLP) with three dense layers to classify audio embeddings into fake or bonafide categories. B. Spectrogram Evaluation As each type of spectrogram presents distinct acoustic fea- ture and ensemble of multiple spectrogram inputs is effective arXiv:2604.19652v1 [cs.SD] 21 Apr 2026 .... Spectrograms 4-second Recording .... Augmented Spectrograms MEL/GAM/CQT CNN-based Backbone MLP Front-end Feature Extraction On-line Augmentation Back-end CNN-based model . . . . Audio Embeddings p fake p real Predicted Probabilities Fig. 1.The high-level architecture of the proposed deep-learning framework Bonafideloss (L3) Backbone .... X .... Backbone .. A-Softmaxloss (L1) Contrastive loss (L2) .. FC –BN –GELU FC –Softmax .. FC –BN –GELU FC –BN –GELU .. .. .. X .. .. .. .. Y Z .. .. Frozen 4-second segments Contrastive Head Softmax Head Backbone .. .. X .... 4-second segments FC –Softmax Entropy loss Entropy Head Transfer Stage 1 Stage 2 Stage 3 Multiple Classes & Multiple Losses Training /wo Mixup(20 epoch) Two-class Finetuning /w Mixup(10 epoch) Two-class Finetuning /wo Mixup(5 epoch) 4-second segments FC –Softmax Entropy loss Entropy Head Fig. 2.The proposed three-stage training strategy to enhance the task performance of Acoustic Scene Classi- fication and Acoustic Event Detection [9], [10], [11], [12], this inspires us to evaluate the role of different spectrograms for the task of ESDD in this paper. Therefore, we conduct experiments on three individual spectrograms of Constant Q Transform (CQT), (MEL), and Gammatone (GAM). We also evaluate if an ensemble of these spectrograms is possible to enhance the ESDD task performance. C. Deep Neural Network Backbone Evaluation To indicate the most effective network architecture, we evaluate four popular network architectures of ResNet50, InceptionV3, EfficientNetB1, and DenseNet161. Addition- ally, we also evaluate the pre-trained BEATs [13] model, which was trained on the large-scale sound event dataset of AudioSet [14]. Regarding the pre-trained BEATs model, we proposed two evaluating methods. In the first method, the pre-trained model is frozen to extract audio embeddings. The audio embedding is then classified into fake and bonafide categories by the MLP in Figure 1. In the second method, all trainable parameters in the pre-trained BEATs and the MLP are finetuned to adapt with the downstream ESDD task. D. The Three-Stage Training Strategy In our previous work [15], we proposed a novel training strategy which proves effective to separate bonafide and fake distribution for deepfake speech detection. This inspire us to apply this training strategy for the ESDD task in this paper. Particularly, the training strategy is modified and described in Fig. 2 with three phases of training processes. In the first phase, we train the model with three loss functions for 20 epoch and the learning rate of 5E-4. The first loss function, A-Softmax loss, is for multiple-class classification (i.e., we consider each fake generator as a category). The second loss, Contrastive loss, is used to separate bonafide and fake distribution. The final loss, Central loss, is used to condense the distribution of bonafide. The combination of three loss is effective to separate fake and bonafide distribution. In the second phase, we train the model with one loss function of Cross Entropy and Mixup data augmentation for 10 epoch and the low learning rate of 1E-5. Finally, we train the model with Cross Entropy without Mixup data augmentation for 5 epoch and the learning rate of 1E-6. Notably, the deep neural network backbone is frozen in the final phase. I. DATASET AND EXPERIMENTAL SETTINGS A. Dataset EnvSDD dataset [3]: In this paper, we train and eval- uate our proposed models on the benchmark dataset of EnvSDD [3]. As shown in Table I, the dataset comprises two subsets of the development subset and the test subset. The bonafide audio in the EnvSDD development subset are from four resources: DCASE Challenge 2019 Task 1 (TU- TASC2019Dev) [16] (e.g., audio scene), DCASE Challenge 2016 Task 3 (TUTSED2016Dev & TUTSED2016Eval) [17] (e.g., audio event), DCASE Challenge 2017 Task 3 (TUTSED2017Dev & TUTSED2017Eval) [18] (e.g., audio event), and finally UrbanSound8K [19] (e.g., audio event). For each data resource, the number of 4-second bonafide audio recordings are collected and shown in the second column of Table I. Then, four different generated-AI systems (e.g., ATA-Audioldm1, TTA-audiogen, TTA-audioldm1, and TTA-audioldm2), using two main techniques of audio-to- audio (ATA) and text-to-audio (TTA), are applied to generate fake audio from bonafide audio. Meanwhile, the bonafide audio in the EnvSDD test subset does not only comprise four resources as mentioned in the development set, but also includes DCASE2023Task7 [20] (e.g., audio event) and Clotho [21] (e.g., audio event). Additionally, three different generated-AI systems of ATA-Audioldm2, TAT-Audiolcm, and TAT-Tangoflux are added to generate fake audio in the test set. Given data in Table I, there are two main concerns. First, the EnvSDD dataset presents unbalanced with the dom- inant TUTASC2019Dev. Second, while TUTASC2019Dev presents sound scene recordings, other data resources col- lect sound events. As acoustic scene and acoustic events show different acoustic features (i.e., Section 1.1.1 in [22] describes the difference between acoustic scene and acoustic event), this raises a question that whether or not we need individual models for different tasks of sound scene fake detection and sound event fake detection. ESDD-Challenge-TestSet [4]: To evaluate the generality of proposed models, we conduct the cross-dataset evaluation with ESDD-Challenge-TestSet. This dataset comprises 1994 4-second bonafide audio and 7981 fake audio. The bonafide audio is from VGG-Sound [23] which presents both sound scene and sound event. The fake audio is generated from two generators, referred to as ‘diff foley’ and ‘foleycrafter’. B. Experimental Settings Proposed Testing Scenario: We follow the paper [3], use the development subset of EnvSDD to train the proposed models and evaluate the proposed models on the test subset of EnvSDD. Given the data splitting, we first propose three test cases. The first test case (Test-Case-1) is for the task of Sound Scene Fake Detection (SSFD) in which only sound scene recordings (e.g., only bonafide and fake audio from TUTASC2019) are used for training and testing. Meanwhile, the second test case (Test-Case-2) is for Sound Event Fake Detection (SEFD). Similarly, only audio recordings contain- ing fake and bonafide audio events (e.g., EnvSDD dataset with the exception of TUTASC2019) are used for training and testing. In the final test case (Test-Case-3) , we cross- test between two tasks of SSFD and SEFD. In particular, the model for SSFD is used to test on SEFD task and and vice versa. In these test cases, we do not apply the proposed training strategy mentioned in Section I-D. Given the experimental results from three test cases, we propose the best model for individual tasks of SSFD and SEFD. The best model is then trained with the proposed three-phase training strategy and conducted the cross-dataset evaluation on ESDD-Challenge-TestSet. Settings: All proposed models in the paper are imple- mented with Pytorch framework. All experiments in this paper are run with GPU Titan 23 GB . Adam algorithm [24] is used for the optimization. Evaluation Metrics: In this paper, we use Accuracy, F1 score, and AuC score as main metrics to compare perofr- mance among models. We also present the confusion matrix on cross-dataset testing. IV. EXPERIMENTAL RESULTS AND DISCUSSION Acoustic Scene Fake Detection (ASFD: Test-Case-1): In this test case, we evaluate deep-learning based models for ASFD task. Therefore, we only train and test with utterances from TUTASC2019Dev set (e.g., sound scene). As Table I shows, GAM spectrogram outperforms MEL and CQT spectrograms when using the same network architec- ture EfficientNetB1. As GAM presents better performance compared with CQT and MEL, we then evaluate different network architectures with GAM spectrogram input. Among network architectures trained from scratch, EfficientNetB1 presents the highest performance compared with the others. However, when we conduct ensemble method, ensemble of multiple spectrograms is more effective to enhance the ASFD performance compared with the ensemble of multiple network architecture. This indicates that different spectro- grams contains distinct features in sound scene. Regarding the approach of leveraging a pre-trained model, we use the pre-trained BEATs model to extract the audio embeddings. Then, the audio embeddings are explored by a MLP network with 3 dense layers. We referred the network as to BEATs- Emb+MLP We achieve best performance on the pre-trained BEATs model compared with the training-from-scratch ap- proach with spectrogram input. Acoustic Event Fake Detection (AEFD: Test-Case-2): In this test case, we evaluate deep-learning based models for AEFD task. Therefore, we train and test with all audio, ex- cepted in TUTASC2019Dev set. Same as Test-Case-1, results in Table I indicate that GAM spectrogram presents the best performance among the evaluating spectrograms. Regarding network architecture trained from scratch, EfficientNetB1 and Inception-V3 are competitive and outperform the others. TABLE I THE NUMBER OF 4-SECOND BONAFIDE/FAKE AUDIO IN ENVSDD DATASET (THE DEVELOPMENT SUBSET IS UPPER AND THE TEST SUBSET IS LOWER) Data ResourceReal AudioFake Audio ATA-Audioldm1TTA-audiogenTTA-audioldm1TTA-audioldm2ATAAudioldm2TTA-AudiolcmTTA-Tangoflux TUTSED2016Dev (Acoustic Event)10481048104810481048NoneNoneNone DevelopmentTUTSED2016Eval (Acoustic Event)478478478478478NoneNoneNone SubsetTUTSED2017Dev (Acoustic Event)12351235123512351235NoneNoneNone TUTSED2017Eval (Acoustic Event)391391391391391NoneNoneNone UrbanSound8K (Acoustic Event)40714071407140714071NoneNoneNone TUTASC2019Dev (Acoustic Scene)31703170317031703170317031703170 TUTSED2016Dev (Acoustic Event)116116116116116116116116 TUTSED2016Eval (Acoustic Event)5353535353535353 TestTUTSED2017Dev (Acoustic Event)137137137137137137137137 SubsetTUTSED2017Eval (Acoustic Event)4343434343434343 UrbanSound8K (Acoustic Event)452452452452452452452452 DCASE2023-Task7 (Acoustic Event)500500500500500500500500 Clotho (Acoustic Event)500500500500500500500500 TABLE I TEST-CASE-1: ACOUSTIC SCENE FAKE DETECTION (ASFD) ON ENVSDD DATASET SpectrogramsModelsAcc.F1AuC MELEfficientNetB10.880.470.69 CQTEfficientNetB10.910.830.96 GAMEfficientNetB10.950.890.99 All Spec. (ensemble)EfficientNetB10.960.900.99 GAMEfficientNetB10.950.890.99 GAMDenseNet1610.920.850.99 GAMInception-V30.910.840.98 GAMResNet500.890.810.96 GAMAll Networks (ensemble)0.930.880.99 Raw AudioBEATs-Emb+MLP TABLE I TEST-CASE-2: ACOUSTIC EVENT FAKE DETECTION (AEFD) ON ENVSDD DATASET SpectrogramsModelsAcc.F1AuC MELEfficientNetB10.360.340.59 CQTEfficientNetB10.690.500.51 GAMEfficientNetB10.740.620.79 All Spec. (ensemble)EfficientNetB10.780.670.81 GAMEfficientNetB10.740.620.79 GAMDenseNet1610.830.670.77 GAMInceptionV30.810.660.78 GAMResNet500.820.660.78 GAMAll Networks (ensemble)0.860.710.82 Raw AudioBEATs-Emb+MLP The ensemble of multiple spectrograms is also more effective than the ensemble of network architectures and leveraging the pre-trained BEATs model achieves the best performance on AEFD task. Cross-Test Between ASFD and AEFD (Test-Case-3): Given models from Test-Case-1 and Test-Case-2, we evaluate the cross-test between two tasks of ASFD and AEFD. In other words, models trained on ASFD (sound scene) is test on AEFD (sound event) and vice versa. As the results are shown in Table IV, it indicates that training a model on ASFD (sound scene) and testing on AEFD (sound event) is not effective. This is similar when we train a model from scratch on AEFD (sound event) and testing on ASFD (sound scene event). However, training on AEFD (sound event) and testing on ASFD (sound scene) using the pre-trained BEATs model (BEATs-Emb-MLP) is very competitive compared with training and testing on ASFD (sound scene). TABLE IV TEST-CASE-3: CROSS-TEST EVALUATION BETWEEN ASFD AND AEFD ON ENVSDD DATASET ModelsTrainTestAccF1AuC GAM+EfficientNetB1ASFD(scene)AEFD(event)0.870.600.61 GAM+InceptionV3ASFD(scene)AEFD(event)0.880.600.66 GAM+ResNet50ASFD(scene)AEFD(event)0.870.590.66 GAM+DenseNet161ASFD(scene)AEFD(event)0.880.600.68 BEATs-Emb+MLPASFD(scene)AEFD(event) GAM+EfficientNetB1AEFD(event)ASFD(scene)0.740.630.84 GAM+InceptionV3AEFD(event)ASFD(scene)0.850.630.78 GAM+ResNet50AEFD(event)ASFD(scene)0.840.630.78 GAM-DenseNet161AEFD(event)ASFD(scene)0.840.630.80 BEATs-Emb+MLPAEFD(event)ASFD(scene) TABLE V EVALUATE BEATS-FINETUNE+MLP WITH THE THREE-STAGE TRAINING STRATEGY ON ENVSDD DATASET ModelsTrainTestAccF1AuC BEATs-Finetune+MLPASFD (scene)ASFD (scene)0.990.990.99 BEATs-Finetune+MLPASFD (scene)AEFD (event)0.950.890.9 BEATs-Finetune+MLPAEFD (event)AEFD (event)0.940.880.98 BEATs-Finetune+MLPAEFD (event)ASFD (scene)0.960.920.99 BEATs-Finetune+MLPASFD&AEFDASFD (scene)0,980.960.99 (scene&event) BEATs-Finetune+MLPASFD&AEFDAEFD (event)0.960,920.99 (scene&event) The experimental results from three test cases above indicates that leveraging a pre-trained models for ESDD task is the most effective approach. By this way, we can train model on AEFD (sound event) and then apply on the inference process for both AEFD (sound event) and ASDF (sound scene). This leads advantages when currently published sound event datasets are larger and more diverse than sound scene. Finetune the pre-trained BEATs model with proposed three-stage training strategy: As experimental results from three test cases shows, leveraging a pre-trained model is more effective than training a model from scratch. We therefore finetune the pre-trained BEATs model using our proposed three-stage training strategy. This model is referred to as BEATs-finetune+MLP. The BEATs-finetune+MLP is not only evaluated on EnvSDD dataset [3] but is also conducted the cross-dataset evaluation on ESDD-Challenge- TestSet [4]. The experimental results in Table V shows TABLE VI EVALUATE BEATS-FINETUNE+MLP WITH THE THREE-STAGE TRAINING STRATEGY ON ESDD-CHALLENGE-TESTSET DATASET TrainTestAccF1AuCEER EnvSDD-ASFD (scene)ESDD-Challenge-TestSet0.750.480.390.568 EnvSDD-AEFD (event)ESDD-Challenge-TestSet0.880.770.920.152 EnvSDD-ASFD& AEFDESDD-Challenge-TestSet0.730.520.440.575 (scene & event) SystemTest SetSeen SDSeen GMTTAATA EnvSDD-ASFD Test 01✓0.0310.034 Test 02✓×0.0380.046 Test 03 ×✓0.1720.209 Test 04 ×0.1530.176 Average–0,0960,116 EnvSDD-AEFD Test 01✓0.0050.005 Test 02✓×0.0600.003 Test 03 ×✓0.0470.029 Test 04 ×0.1390.015 Average–0.0610.052 EnvSDD-AEFD+ASFD Test 01✓0.0010.001 Test 02✓×0.0030.001 Test 03 ×✓0.0360.031 Test 04 ×0.0430.032 Average–0.0210.016 TABLE VII EER COMPARISON that the finetuning BEATs model is effective to enhance the ESDD performance compared with the embedding-based approach (BEATs-Emb+MLP). We achieve the best model performance with Acc. of 0.98, F1 Score of 0.95, and AuC score of 0.99 on the Test subset of EnvSDD dataset (i.e., these results are computed on both sound event and sound scene of Test subset). Given the high performance of the proposed BEATs- Finetune+MLP on EnvSDD dataset, we evaluate the cross- dataset testing on ESDD-Challenge-TestSet [23]. As Ta- ble VI shows, it indicates that the model with training on sound event from EnvSDD dataset outperforms models with training on sound scene or both sound event and sound scene. In other words, training on sound scenes from EnvSDD dataset leads to downgrade the model performance. This can be explained that audio data in ESDD-Challenge-TestSet is from VGG-Sound [23] which mainly present sound events. V. CONCLUSION We have presented a deep-learning based framework for environmental sound deepfake detection (ESDD). By con- ducting extensive experiments on the benchmark datasets of EnvSDD and ESDD-Challenge-TestSet, we indicate that (1) Detecting deepfake audio of sound event and sound scene should be considered as individual tasks; (2) A model for ESDD on sound event can be effectively apply for ESDD on sound scene; (3) finetuning a pre-trained model is effective for ESDD task rather than training from scratch. We eventually propose the best model for the ESDD task by combining finetuning the pre-trained BEATs model with our proposed three-stage training strategy. REFERENCES [1] “Foleysoundsynthesis,” https://dcase.community/ challenge2023/task-foley-sound-synthesis. [2] Jiangyan Yi, Chenglong Wang, Jianhua Tao, Chu Yuan Zhang, Cun- hang Fan, Zhengkun Tian, Haoxin Ma, and Ruibo Fu, “Scenefake: An initial dataset and benchmarks for scene fake audio detection,” Pattern Recognition, vol. 152, p. 110468, 2024. [3] Han Yin, Yang Xiao, Rohan Kumar Das, Jisheng Bai, Haohe Liu, Wenwu Wang, and Mark D Plumbley,“Envsdd: Benchmarking environmental sound deepfake detection,” in Proc. INTERSPEECH, 2025, p. 201–205. [4] “Esddchallengeinicassp,” https://github.com/ apple-yinhan/EnvSDD, Accessed: 2010-09-30. [5] Hafsa Ouajdi, Oussama Hadder, Modan Tailleur, Mathieu Lagrange, and Laurie M Heller, “Detection of deepfake environmental audio,” in Proc. EUSIPCO, 2024, p. 196–200. [6] Orchid Chetia Phukan et al., “Representation loss minimization with randomized selection strategy for efficient environmental fake audio detection,” arXiv preprint arXiv:2409.15767, 2024. [7] Kele Xu, Dawei Feng, Haibo Mi, Boqing Zhu, Dezhi Wang, Lilun Zhang, Hengxing Cai, and Shuwen Liu, “Mixup-based acoustic scene classification using multi-channel convolutional neural network,” in Pacific Rim Conference on Multimedia, 2018, p. 14–23. [8] Yuji Tokozume, Yoshitaka Ushiku, and Tatsuya Harada, “Learning from between-class examples for deep sound recognition,” in ICLR, 2018. [9] Lam Pham, Lang Yue, et al., “Bag-of-features models based on c-dnn network for acoustic scene classification,” in Proc. AES, 2019. [10] Lam Pham, Dat Ngo, Dusan Salovic, Anahid Jalali, Alexander Schindler, Phu X Nguyen, Khoa Tran, and Hai Canh Vu, “Lightweight deep neural networks for acoustic scene classification and an effective visualization for presenting sound scene contexts,” Applied Acoustics, vol. 211, p. 109489, 2023. [11] Lam Pham, Khoa Tran, Dat Ngo, Hieu Tang, Son Phan, and Alexander Schindler, “Wider or deeper neural network architecture for acoustic scene classification with mismatched recording devices,” in Proceed- ings of the 4th ACM International Conference on Multimedia in Asia, 2022, p. 1–5. [12] Huy Phan, Huy Le Nguyen, Oliver Y Ch ́ en, Lam Pham, Philipp Koch, Ian McLoughlin, and Alfred Mertins, “Multi-view audio and music classification,” in Proc. ICASSP, 2021, p. 611–615. [13] Sanyuan Chen et al.,“BEATs: Audio pre-training with acoustic tokenizers,” in Proceedings of the 40th International Conference on Machine Learning, 2023, vol. 202, p. 5178–5193. [14] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in Proc. ICASSP, 2017. [15] Lam Pham, Dat Tran, Phat Lam, Florian Skopik, Alexander Schindler, Silvia Poletti, David Fischinger, and Martin Boyer,“Din-cts: Low-complexity depthwise-inception neural network with contrastive training strategy for deepfake speech detection,”arXiv preprint arXiv:2502.20225, 2025. [16] “Dcase-2019-challenge-task-1,” https:// dcase.community/challenge2019/ task-acoustic-scene-classification,Accessed: 2010-09-30. [17] “Dcase-2016-challenge-task-3,” https:// dcase.community/challenge2016/ task-sound-event-detection-in-real-life-audio, Accessed: 2010-09-30. [18] “Dcase-2017-challenge-task-3,” https:// dcase.community/challenge2017/ task-sound-event-detection-in-real-life-audio, Accessed: 2010-09-30. [19] J. Salamon, C. Jacoby, and J. P. Bello, “A dataset and taxonomy for urban sound research,” in Proc. ACM-M, 2014, p. 1041–1044. [20] “Dcase-2023-challenge-task-7,” https://dcase.community/ challenge2023, Accessed: 2010-09-30. [21] Konstantinos Drossos, Samuel Lipping, and Tuomas Virtanen, “Clotho: an audio captioning dataset,” in Proc. ICASSP, 2020, p. 736–740. [22] Lam Dang Pham,Robust deep learning frameworks for acoustic scene and respiratory sound classification, University of Kent (United Kingdom), 2021. [23] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman, “Vggsound: A large-scale audio-visual dataset,”in International Conference on Acoustics, Speech, and Signal Processing, 2020. [24] P. K. Diederik and B. Jimmy,“Adam: A method for stochastic optimization,” CoRR, vol. abs/1412.6980, 2015.