Multiband Dynamics Compressor with Automatic Parameter Learning

Dynamic range compression (DRC) is vital in music production, mixing, mastering, TV broadcasting, and film post-processing. It evens out sound by reducing the difference between the loudest and softest parts of an audio signal, balancing levels, preventing distortion, and enhancing quality. Recognizing its importance, audio engineers rely heavily on DRC. By splitting the frequency spectrum into multiple bands, traditional compressors gain even more precision. This paper introduces a four-band dynamics compressor developed with the JUCE framework, overcoming significant challenges in implementing Linkwitz-Riley filters for more than three bands, while setting the stage for advanced machine learning algorithms in audio engineering. The primary focus, thus, is on a machine learning model based on a siamese neural network that automatically learns the parameters of the multi-band dynamic range compressor. This model can capture the style of an experienced audio engineer by analyzing features from compressed and uncompressed audio spectrograms. The goal is to create a tool that learns and adapts to professional users’ specific styles, even with smaller datasets, and predicts all the necessary parameters of the compressor for new audio files.