On the evening of November 20, SCS's 3rd @ World "Touching Academic Frontiers" international exchange event in 2023 was successfully held in the conference room at the corner of the 2nd floor. This event invited six outstanding students from the SCS, Ning Ziqian, Zhang Zihan, Wang Zhichao, Song Kun, Hu Shishuai, and Zhang Shaoteng, to share their latest research findings and applications in the fields of intelligent speech processing and medical imaging intelligent computing. They elaborated on the cutting-edge issues and challenges in their respective research fields in simple and concise language, bringing a rich academic feast for the students present. Such face-to-face academic exchange not only deepens students' understanding of different fields, but also stimulates their interest and enthusiasm for academic research.
Ning Ziqian shared his ICASSP participation experience and gave a brief introduction to his paper on high performance speech conversion. High performance speech conversion is a challenging research direction. But current solutions have to balance speaker similarity, comprehensibility, and expressiveness. To address this issue, Ning and his team proposed an end-to-end speech conversion model, Expressive-VC, which combines the advantages of bottleneck feature based (BNF) method and information perturbation based method. Experiments have shown that Expressive-VC is superior to multiple comparison systems, as it can capture high expressiveness from the source speech, and achieve high similarity with the target speaker while maintaining good intelligibility.
Zhang Zihan brought a sharing about the AEC Challenge participation system, the invited paper, and his ICASSP 2023 participation experience. At the AEC Challenge, Zhang proposed a two-step frequency division band neural network (TBNN) for full band residual echo cancellation. The main principle is to divide the full band signal into two parts: broadband (16 KHz) and high-frequency (16-48 KHz) after linear filtering, and model it with lower complexity while suppressing residual echoes and noise. Broadband signals are processed by a gated convolutional recursive network using a U ^ 2 encoder, while high-frequency signals are processed by a lower complexity high-frequency post filtering network. This plan ultimately achieved an overall average opinion score (MOS) of 4.344 and a word accuracy rate (WAcc) of 0.795, ranking second (tied) in the non personalized track.
Wang Zhichao gave a report titled "Introduction to ICASSP2023 Attending Experience and Paper on Low Resource Voice Conversion". In voice conversion (VC), it is crucial to convey language content to the target speaker while maintaining the speaking style of the source speech. However, in low resource scenarios where the target speaker's corpus is highly limited, existing VC methods can hardly meet this requirement. In response to this issue, Wang and his team proposed a new VC model for handling low resource VC tasks. In addition, Wang and his team also introduced a simulation mode during training to simulate the inference process, alleviating the mismatch between training and inference. Experiments on high performance voice demonstrate that the proposed method performs well on low resource VC tasks.
Song Kun shared his experience of attending Interspeech2023 and gave a brief introduction to the team's paper on speech to speech translation. Hu Shishuai shared his MICCAI2022 attending experience and his paper on unsupervised domain adaptive medical image segmentation reconstructed based on domain specific convolution and high-frequency component with his classmates. Zhang Shaoteng briefed on the team's achievements in weakly supervised tissue pathology image segmentation based on Transformer.
The @ World "Touching Academic Frontiers" international exchange event provides students with a platform to learn about the latest research trends and academic frontiers in related fields, broaden their international academic horizons, and increase their professional knowledge. We believe that through such a platform, students can continuously improve their professional competence and comprehensive abilities, laying a solid foundation for future academic research and career development. It is hoped that the @ World "Touching Academic Frontiers" international exchange series activities will become an important milestone and new driving force in everyone's academic exchange activities!
