Investigating the 'Autoencoder Behavior' in Speech Self-Supervised Models: a focus on HuBERT's Pretraining
This addresses a performance bottleneck in speech recognition models for researchers and practitioners, but it is incremental as it builds on existing HuBERT insights.
The paper investigates the 'autoencoder behavior' in speech self-supervised models, where finetuning all layers reduces performance due to top layers retaining input-like information, and finds that improving HuBERT's training procedure leads to faster convergence and competitive results on downstream tasks.
Self-supervised learning has shown great success in Speech Recognition. However, it has been observed that finetuning all layers of the learned model leads to lower performance compared to resetting top layers. This phenomenon is attributed to the ''autoencoder'' behavior: top layers contain information closer to the input and are less suitable for tasks that require linguistic information, such as Speech Recognition.To better our understanding of this behavior, we propose to study the evolution of high-level information within the model during pretraining. We focus on the HuBERT model, which exhibits a less pronounced ''autoencoder'' behavior. By experimentally exploring various factors that may have an impact, we aim to improve the training procedure and enhance the top layers of HuBERT for high-level tasks.Furthermore, our experiments demonstrate that these improvements in the training procedure result in faster convergence and competitive performance on downstream tasks.