New metric for audio compressors beats waveform-based evaluation

A new paper from Benjamin R. Thompson and Michael C. Heilemann tackles a fundamental problem in audio machine learning: how to accurately evaluate dynamic range compressor models. Current approaches rely on waveform-based proxy metrics from audio input-output data, but isolating the true gain reduction signal creates an ill-conditioned inverse problem. These proxy metrics can be influenced by inaudible analog processing artifacts, making it unclear how well they reflect the actual compressor behavior. The authors propose a direct measurement approach—using the hardware's control voltage signal as ground truth for gain reduction, bypassing audio waveform distortion entirely.

To validate their method, Thompson and Heilemann trained a gray-box model using loss computed directly over the gain control signals, alongside two models trained with common proxy losses. The proxy-trained models failed to match the performance of the direct-control model when evaluated against the underlying control trajectory. Waveform-domain metrics assigned similar errors to models that were clearly distinguished by the direct metric, exposing a significant blind spot. The researchers also released a dataset of a Solid State Logic bus compressor that includes synchronized gain control voltage and audio output, enabling further exploration of this evaluation paradigm. This work could reshape how audio effects are modeled, offering a more accurate and robust evaluation standard for emulating analog hardware.