Dilated Residual Network automates resonance equalization for audio production

In music and audio production, attenuating spectral resonances—sharp frequency peaks caused by recording environments or instruments—is critical for clean mixes. Maarten Grachten, Emmanuel Deruty, and Alexandre Tanguy propose an auto-adaptive system to automate this tedious task. Their system combines a dynamic equalizer that automatically detects resonances and offers user-specified attenuation, with a deep neural network that predicts the optimal attenuation factor from windowed audio segments. The network is trained on empirical data from an experiment where sound engineers selected their preferred attenuation factors for a set of tracks.

The team tested two neural architectures for the predictive model: a dilated residual network operating directly on the raw audio waveform, and a baseline network requiring a separate audio feature extraction stage. Remarkably, the dilated residual network performed on par with the feature-based approach, both significantly beating a simple baseline. This shows that end-to-end learning from raw audio can match handcrafted feature methods for this task, simplifying deployment and potential real-time use. Results were presented at ISMIR 2019, highlighting the potential for AI-assisted audio processing tools that learn from expert human judgment.