On Fine-Tuning Pre-Trained Speech Models With EMA-Target Self-Supervised Loss

International Conference
2023-12-14 16:35
Authors : Hejung Yang, Hong-Goo Kang

Year : 2024

Publisher / Conference : ICASSP

Research area : Speech Signal Processing

Presentation/Publication date : 2024.04.19

Presentation : Poster

Representation models pre-trained on self-supervised objectives are often fine-tuned for solving downstream tasks.
However, fine-tuning can degrade the general knowledge that was originally built up by the pre-training, which could help prevent the model from overfitting given sparse fine-tuning data or bridge gaps between different domains.
We hypothesize that preserving this general knowledge in pre-trained models is crucial for improving performance on downstream tasks.
Based on this idea, we propose a novel method for fine-tuning self-supervised speech models that utilizes a self-supervised loss over the course of fine-tuning.
Then, an Exponential Moving Average (EMA) technique is applied to smoothly transition the domain of the model from the generalized to the task-oriented one.
We perform various downstream tasks using the proposed method, finding that our method improves performance on most of the tasks. Results show that our method induces the generalization ability of the model to be retained without overshadowing the downstream task performance.
전체 355
29 International Conference Hyewon Han, Naveen Kumar "A cross-talk robust multichannel VAD model for multiparty agent interactions trained using synthetic re-recordings" in Hands-free Speech Communication and Microphone Arrays (HSCMA, Satellite workshop in ICASSP), 2024
28 International Conference Yanjue Song, Doyeon Kim, Nilesh Madhu, Hong-Goo Kang "On the Disentanglement and Robustness of Self-Supervised Speech Representations" in International Conference on Electronics, Information, and Communication (ICEIC) (*awarded Best Paper), 2024
27 International Conference Yeona Hong, Miseul Kim, Woo-Jin Chung, Hong-Goo Kang "Contextual Learning for Missing Speech Automatic Speech Recognition" in International Conference on Electronics, Information, and Communication (ICEIC), 2024
26 International Conference Juhwan Yoon, Seyun Um, Woo-Jin Chung, Hong-Goo Kang "SC-ERM: Speaker-Centric Learning for Speech Emotion Recognition" in International Conference on Electronics, Information, and Communication (ICEIC), 2024
25 International Conference Hejung Yang, Hong-Goo Kang "On Fine-Tuning Pre-Trained Speech Models With EMA-Target Self-Supervised Loss" in ICASSP, 2024
24 International Conference Hong-Goo Kang, W. Bastiaan Kleijn, Jan Skoglund, Michael Chinen "Convolutional Transformer for Neural Speech Coding" in Audio Engineering Society Convention, 2023
23 International Conference Hong-Goo Kang, Jan Skoglund, W. Bastiaan Kleijn, Andrew Storus, Hengchin Yeh "A High-Rate Extension to Soundstream" in IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2023
22 International Conference Zhenyu Piao, Hyungseob Lim, Miseul Kim, Hong-goo Kang "PDF-NET: Pitch-adaptive Dynamic Filter Network for Intra-gender Speaker Verification" in APSIPA ASC, 2023
21 International Conference Miseul Kim, Zhenyu Piao, Jihyun Lee, Hong-Goo Kang "BrainTalker: Low-Resource Brain-to-Speech Synthesis with Transfer Learning using Wav2Vec 2.0" in The IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI), 2023
20 International Conference Seyun Um, Jihyun Kim, Jihyun Lee, Hong-Goo Kang "Facetron: A Multi-speaker Face-to-Speech Model based on Cross-Modal Latent Representations" in EUSIPCO, 2023