딥 보이스 탐지 성능향상을 위한오인률 기반 가중치 앙상블 기법
Error Rate-Based Weighted Ensemble Method for Improving the Performance of Deep Voice Det
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
최근 AI 기반의 음성 합성 기술의 발전으로 딥 보이스와 같은 합성 음성을 이용한 보이스 피싱과 같은 범죄가 증가하고 있다. 기존의 딥 보이스 탐지 모델들은 Accuracy 기반 가중치 앙상블 기법을 사용하지만, 오인률을 충분히 반영하지 못하는 한계가 있다. 본 논문에서는 딥 보이스 오인률(Deep Voice Error Rate, DE)과 실제 음성 오인률(Real Voice Error Rate, RE)의 차이를 반영하여 각 모델의 오인률을 가중치로 부여하는 새로운 앙상블 기법을 제안한다. 실험을 통해 제안 기법이 기존 기법보다 높은 딥 보이스 탐지 정확도와 낮은 DE와 RE를 보였다. 또한, 훈련 데이터와 테스트 데이터는 다른 음성 합성 기술로 변환된 데이터를 사용함으로써 제안된 앙상블 기법이 새로운 음성 합성 기술에 적용 가능할 것으로 기대된다.
Abstract
Recent advances in AI-based speech synthesis have led to an increase in crimes, such as voice phishing, usingsynthetic voices like Deep voice. Existing deep voice detection models rely on accuracy-based weighted ensembletechniques, but they fail to fully reflect error rates. This paper proposes a new ensemble method that weightsmodels based on the difference between Deep Voice Error Rate (DE) and Real Voice Error Rate (RE). Experimentsshow that the proposed method achieves higher deep voice detection accuracy and lower DE and RE compared toexisting methods. Additionally, the training data and test data are converted from different speech synthesistechniques, which is expected to make the proposed ensemble method applicable to new speech synthesistechniques.
참고문헌 (11)
- Arık, Sercan Ö, "Deep voice: Real-time neural text-to-speech", Proceedings of the 34th International conference on machine learning, pp.195~204 (2017), PMLR
- Sowoon Kim, "Development of Voice Phishing Damage Prevention Service Misusing Deep Voice", The Journal of Korean Institute of Communications and Information Sciences, 47(10), pp.1677~1685 (2022)[DOI]
- "Crying daughter in video turns out to be ‘deepfake’... Kidnapping situation? Report it immediately"[URL]
- Pham, Lam, "Deepfake Audio Detection Using Spectrogram-based Feature and Ensemble of Deep Learning Models", Proceedings of the 2024 IEEE 5th International Symposium on the Internet of Sounds (IS2), pp.1~5 (2024)
- T. Gopalakrishnan, "Proceedings of the 2022 IEEE International Conference on Data Science and Information System (ICDSIS)", Proceedings of the 2022 IEEE International Conference on Data Science and Information System(ICDSIS), pp, pp.1~6 (2022)
- Meghanani, Amit, "An exploration of log-mel spectrogram and MFCC features for Alzheimer’s dementia recognition from spontaneous speech", Proceedings of the 2021 IEEE spoken language technology workshop (SLT), pp.670~677 (2021)
- Jadeja, Viswajeetsinh, "Convolutional Neural Networks: A Comprehensive Review of Architectures and Application", Proceedings of the 2023 6th International Conference on Contemporary Computing a nd Informatics (IC3I), pp.460~467 (2023)
- Alnuaim, Abeer Ali, "Speaker gender recognition based on deep neural networks and ResNet50", Wireless Communications and Mobile Computing, 2022(1), pp.1~13 (2022)[DOI]
- Tiwari, Manish, "Enhanced text-independent speaker recognition using MFCC, Bi-LSTM, and CNN-based noise removal techniques", International Journal of Speech Technology, 27, pp.1013~1026 (2024)[DOI]
- Wu, Minglin, "Transformer Based Endto-End Mispronunciation Detection and Diagnosis", Proceedings of the Interspeech, pp.3954~3958 (2021)