상세정보
학술대회
Multi-Scale Multi-Band Dilated DenseLSTM for Robust Recognition of Speech with Background Music
- 저자
- 허운행, 김혜미, 권오욱
- 발행일
- 202010
- 출처
- International Conference on Information and Communication Technology Convergence (ICTC) 2020, pp.1238-1241
- 협약과제
- 20IH2400, 음악 및 동영상 모니터링을 위한 지능형 마이크로 식별 기술 개발, 박지현
- 초록
- We propose a multi-scale multi-band dilated time- frequency DenseNet with LSTM for speech enhancement and speech recognition. In the convolutional neural network (CNN)- based architecture, it is important to increase the receptive field effectively in order to sufficiently consider the context information. In our previous study, we designed a dilated dense block that reflects acoustic characteristics by applying dilated convolutions to a densely connected convolutional network (DenseNet) in order to effectively increase the receptive field. In this study, for speech enhancement, we apply the dilated dense blocks to MMDenseLSTM based on a convolutional recurrent neural network (CRNN) which has shown good performance in recent studies using a deep learning architecture. We conduct a speech enhancement and speech recognition experiment using the proposed architecture and several existing deep learning architectures: Gated residual network (GRN), MMDenseLSTM, DilDenseNet. Overall, the proposed architecture shows the best performance compared to other deep learning architectures.
- KSP 제안 키워드
- Acoustic characteristics, Background music, Best performance, Context Information, Convolution neural network(CNN), Convolutional networks, Deep Learning Architectures, Dense block, Multi-scale, Receptive field, Recurrent Neural Network(RNN)