Tackling Long-Range Dependencies in Dynamic Range Compression Modeling via Deep Learning

This chapter delves into the application of deep learning techniques to model dynamic range compression (DRC) in audio effects, with a specific focus on the API-2500+ Stereo Bus Compressor. The study introduces a novel architecture called Series-Parallel Temporal Modulation (SPTMod), which is designed to efficiently handle long duration time dependencies. The research compares SPTMod with state-of-the-art models like GCNTF and TCN, evaluating their performance using various metrics such as L1, ESR, MR-EESR, and MR-STFT. A key innovation in this study is the introduction of the state prediction network, which initializes the states of recurrent layers, reducing the need for warm-up and lowering computational costs. The chapter also explores the impact of different training strategies, including windowed target and streamed target approaches, and discusses the implications of exposure bias in models using state prediction. The results indicate that SPTMod achieves similar performance to GCNTF but with better spectral accuracy, highlighting its potential for improving audio effect modeling. The study concludes by suggesting future work to address the discrepancy in streamed target error and to explore the application of state prediction to other audio effects.