How to Estimate Model Transferability of Pre-Trained Speech Models?
Journal
Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH
Journal Volume
2023-August
Date Issued
2023-01-01
Author(s)
Chen, Zih Ching
Yang, Chao Han Huck
Li, Bo
Zhang, Yu
Chen, Nanxin
Chang, Shou Yiin
Prabhavalkar, Rohit
Sainath, Tara N.
Abstract
In this work, we introduce a “score-based assessment” framework for estimating the transferability of pre-trained speech models (PSMs) for fine-tuning target tasks. We leverage upon two representation theories, Bayesian likelihood estimation and optimal transport, to generate rank scores for the PSM candidates using the extracted representations. Our framework efficiently computes transferability scores without actual fine-tuning of candidate models or layers by making a temporal independent hypothesis. We evaluate some popular supervised speech models (e.g., Conformer RNN-Transducer) and self-supervised speech models (e.g., HuBERT) in cross-layer and cross-model settings using public data. Experimental results show a high Spearman's rank correlation and low p-value between our estimation framework and fine-tuning ground truth. Our proposed transferability framework requires less computational time and resources, making it a resource-saving and time-efficient approach for tuning speech foundation models.
Subjects
and foundation speech models | model transferability | Pre-trained speech models | transfer learning
Type
conference paper