We won a best paper award at the ICEIC conference (International Conference on Electronics, Information and Communication) for our paper "On the Disentanglement and Robustness of Self-Supervised Speech Representations"
Authors: Yanjue Song (UGent), Doyeon Kim (Yonsei University, South Korea), Nilesh Madhu (UGent), Hong-Goo Kang (Yonsei University, South Korea)
Short info: This paper evaluates latent embeddings provided by several pre-trained networks (HuBERT, TERA, wav2vec and wavLM) with respect to their sensitivity to noise and reverberation. We show that speaker-related information is not equally distinguishable for each network. While TERA and wavLM seem most robust to distortion caused by noise and reverberation, TERA seems to provide the most speaker-related discrimination while HuBERT and wavLM provide embeddings that are more sensitive to the phonetic context. These results indicate that the selection of the model should be based on the specific downstream task.
This is joint work with the DSP&AI Lab of Prof. HONG-GOO KANG at Yonsei University.
Thank you, Yanjue Song and Doyeon Kim for this thought-provoking study.
