Audio Active Learning With Noisy Labels

Abstract

Audio annotation is particularly costly and prone to errors due to the temporal nature and semantic ambiguity in audio perception. Active learning (AL) addresses this by iteratively selecting the most informative samples from an unlabeled pool for expert labeling, thereby maximizing model performance with minimal annotation effort. In the standard continual fine-tuning framework widely adopted in audio AL, the model trained in each cycle serves as the initialization for the next, preserving accumulated knowledge to enhance both sample selection and model accuracy. However, we discover a critical limitation when annotation noise is inevitably introduced: this default continual fine-tuning approach becomes susceptible to Primacy Bias — a phenomenon where early-learned patterns persistently influence subsequent learning. Our experiments show that this bias causes the model to overfit noise more rapidly in later cycles. To our knowledge, this represents the first comprehensive study specifically addressing Active Learning with Noisy Labels (ALNL) in the audio domain. To address this issue, we introduce a re-initialization strategy for ALNL scenarios. Our experiments demonstrate that periodically resetting model parameters preserves the model’s ability to learn from clean samples. Furthermore, we propose the Self-Purify Active Learning (SPAL) method, which dynamically identifies potential label noise via training loss modeling and supports either human-in-the-loop correction or automated label refurbishment. Extensive experiments on underwater acoustics, general audio, and speech datasets demonstrate the effectiveness of our framework against label noise in AL scenarios.