Reverse engineering of a recording mix with differentiable digital signal processing

Area of Science:

• Audio Signal Processing
• Music Information Retrieval
• Computational Acoustics

Background:

• Audio mixing involves complex signal processing chains.
• Recovering mixing parameters from a final mix is challenging.
• Existing methods often use blackbox models, lacking interpretability.

Purpose of the Study:

• To develop a method for retrieving audio mixing parameters from raw tracks and stereo mixdown.
• To model linear time-invariant audio effects, including equalization, delay, and reverb.
• To propose and evaluate interpretable audio effect models for music production.

Main Methods:

• Utilized differentiable digital signal processing modules and stochastic gradient descent for optimization.
• Developed two reverb module architectures: stereo reverb and individual reverb.
• Compared model performance using objective feature measures and a perceptual listening study.

Main Results:

• The stereo reverb model demonstrated superior performance on objective measures.
• No statistically significant difference was found between the perception of the stereo reverb model and reference mixes.
• The method successfully modeled linear time-invariant effects like gain, pan, equalization, delay, and reverb.

Conclusions:

• The proposed method provides an interpretable representation of the audio mixing signal chain.
• The stereo reverb model is effective for parameter retrieval and perceptual mix reconstruction.
• This approach advances the understanding and automation of audio mixing processes.