Audio Effect Estimation with DNN-Based Prediction and Search Algorithm

Researchers Youichi Okita and Haruhiro Katayose have introduced a novel method for audio effect estimation, presented in a paper accepted at ICASSP 2026. Their approach combines DNN-based prediction with a search algorithm to reverse-engineer audio effects from a wet signal. First, the DNN predicts the dry signal and effect configuration, then a search refines these predictions using wet signal reconstruction similarity as an objective function. This hybrid method outperforms traditional predictive-only approaches, with the most effective strategy being to predict effect type combinations first, followed by search-based estimation of order and parameters.

This work addresses a key challenge in audio processing: accurately estimating the configuration of applied effects from processed audio. By integrating predictive and search-based methods, the researchers achieve superior performance across various metrics. The paper, submitted to arXiv (ID: 2604.22276), demonstrates that task division—predicting effect types then searching for order and parameters—yields the best results, offering a practical solution for sound designers and audio engineers. The findings suggest that this hybrid approach can significantly improve audio effect estimation accuracy, with potential applications in music production, audio restoration, and sound design automation.