SELF-TAILORING PROMPTS FOR PARAMETER EFFICIENT TUNING SPEECH RECOGNITION

Dianwen Ng, Chong Zhang, Ruixi Zhang, Yukun Ma, Trung Hieu Nguyen, Chongjia Ni, Eng Siong Chng, Bin Ma

Keywords: Parameter Efficient Fine-tuning, Soft Prompt Tuning, Automatic Speech Recognition

Abstract: Soft-prompt tuning is an emerging topic for speech recognition despite its success in many natural language processing tasks. Although it appears to be a promising approach for efficiently fine-tuning large speech models, it can suffer from subpar prompt generalization and a lack of instance-specific guidance due to its "one-size-fits-all" template. To address these limitations, we propose a self-tailoring prompting mechanism that adaptively modifies the prompt tokens to incorporate relevant speech utterance-specific information. Self-tailoring mechanism includes simple yet effective prompt masking regularization techniques and a redundancy reduction loss to improve the quality of soft prompt tokens. Extensive experiments demonstrate that our method achieves better generalization capability and consistently achieves improved performance on the speech recognition task under a wide range of acoustic scenarios, including both clean and noisy speech environments. Self-tailoring prompt tuning outperforms the full fine-tuning model with as few as 0.7% of its trainable weights.

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2469