⚡ electrical engineering

Evaluating Dynamic Range Compressor Models Using Control-Voltage Measurements: an Approach and Dataset

This paper proposes a novel evaluation method for dynamic range compressor models that directly compares predicted gain-reduction signals against hardware control-voltage measurements, demonstrating its superiority over traditional waveform-based proxy metrics and releasing a new dataset containing these control signals to facilitate further research.

Original authors: Benjamin R. Thompson, Michael C. Heilemann

Published 2026-06-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Benjamin R. Thompson, Michael C. Heilemann

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to mimic a famous, vintage sound compressor—a device used by music producers to control the volume of a song, making loud parts quieter and quiet parts louder. The goal is to get the robot to behave exactly like the real machine.

The problem, according to this paper, is that we've been grading the robot's homework using the wrong ruler.

The Wrong Ruler: Listening to the Result

Usually, to see if the robot is doing a good job, engineers play a song into the real machine and into the robot, then compare the two resulting songs. They look at the final audio waves and ask, "How different do these sound?"

The authors argue this is like judging a chef's ability to bake a cake by tasting the final cake, without knowing if the chef actually followed the recipe. The real machine might add a tiny bit of "noise" or change the sound slightly in ways that aren't musical but are just side effects of its old electronics. If the robot copies those weird side effects, it might get a high score on the "taste test" (the audio comparison) even if it completely misunderstood the recipe (the volume control logic). Conversely, if the robot follows the recipe perfectly but misses a tiny, inaudible electronic quirk, it might get a low score.

The paper calls these "proxy metrics." They are indirect measurements that get confused by the "noise" of the recording process.

The Right Ruler: Watching the Hand

The authors propose a better way: instead of just listening to the final song, we should watch the robot's "hand" as it turns the volume knob.

In the real hardware machine, there is a specific electrical signal (a control voltage) that tells the volume knob exactly how much to turn up or down at every single moment. This is the "recipe" in action.

The researchers built a new dataset where they didn't just record the music going in and out of the machine; they also recorded that specific "volume knob signal" coming out of the machine. This allows them to compare the robot's internal "volume knob signal" directly against the real machine's signal.

The Experiment: Training with the Right Ruler

To prove this new method works, they ran a test:

They trained three different robot models to mimic the machine.
Robot A was trained using the old method (comparing the final songs).
Robot B was trained using a slightly better version of the old method (comparing energy levels).
Robot C was trained using the new method (comparing the "volume knob signals" directly).

The Result: When they checked all three robots against the actual volume knob signal of the real machine, Robot C was the clear winner. Robots A and B thought they were doing well because their final songs sounded similar, but their internal logic for controlling the volume was actually quite different from the real machine. The old methods were "hiding" the mistakes by focusing on the wrong part of the process.

The New Dataset: A Goldmine for Researchers

Finally, the authors released a massive new collection of data (about 270 GB) to help others. It includes:

Music: 219 different 30-second song clips.
The Machine: A specific, famous SSL bus compressor (a classic piece of studio gear).
The Secret Sauce: For every song, they recorded the audio and the exact "volume knob signal" that the machine used.

They also included special "test tones" (like a slowly rising volume ramp) to help researchers figure out exactly how the machine reacts to different settings.

The Bottom Line

This paper says that if you want to build a digital model that truly understands how a compressor works, you shouldn't just listen to the output. You need to measure the control signal directly. They have provided the tools (the dataset) and the proof (the experiment) to show that this direct measurement is the only way to get a truly accurate model.

Technical Summary: Evaluating Dynamic Range Compressor Models Using Control-Voltage Measurements

Problem Statement
The primary behavior of a dynamic range compressor (DRC) is defined by the time-varying gain signal ( $g_{dB}[k]$ ) applied to an input signal. However, in existing datasets for DRC modeling, this gain signal is not directly observable; only audio input–output pairs are provided. Isolating $g_{dB}[k]$ from audio data constitutes an ill-conditioned inverse problem, complicated by secondary effects inherent in analog hardware and capture chains, such as phase shifts, amplitude-response deviations, and noise. Consequently, current model evaluation relies on "proxy metrics" computed over the output waveforms (e.g., Mean Absolute Error or Multi-Resolution Short-Time Energy). The authors argue that these proxy metrics are sensitive to the same secondary effects that obscure the gain signal, potentially leading to misleading evaluations where models are penalized or rewarded for mimicking inaudible artifacts rather than the core gain-reduction behavior.

Methodology
To address these limitations, the authors propose a direct evaluation method using control-voltage (CV) measurements. The core methodology involves:

Direct Measurement: Capturing the gain-reduction control voltage signal produced by hardware alongside the audio input and output. This CV signal is scaled to represent the exact gain reduction in decibels ( $g_{dB}[k]$ ), bypassing the need for inverse problem solving.
Experimental Comparison: The authors trained a gray-box dynamics model (a modified version of torchcomp) using three distinct loss regimes:
- Direct loss on the gain signal: $L_1(g_{dB}, \hat{g}_{dB})$ .
- Proxy loss on processed waveforms: $L_1(y, \hat{y})$ .
- Proxy loss on processed waveforms using Multi-Resolution Short-Time Energy: $MSTE(y, \hat{y})$ .
Evaluation: Ten independent models were trained for each regime using identical optimization settings and a single 30-second stereo excerpt. The models were then evaluated against the ground-truth gain trajectory and the proxy metrics to determine if minimizing proxy losses effectively recovers the underlying control trajectory.

Key Contributions

Methodological Approach: The paper introduces and validates a method for evaluating DRC models by measuring the gain-reduction signal directly against the hardware control voltage, rather than inferring it from audio waveforms.
Dataset Release: The authors present a new dataset featuring a Solid State Logic (SSL) Logic FX G384 bus compressor. It includes 219 stereo music clips and calibration signals processed across 90 control permutations. Crucially, every audio input–output pair is accompanied by the corresponding gain control voltage signal ( $g_{dB}[k]$ ), sampled at 44.1 kHz with 32-bit resolution.
Empirical Evidence: The study provides empirical data demonstrating that proxy metrics (waveform-based $L_1$ and MSTE) fail to distinguish between models that have significantly different gain-reduction accuracies.

Results
The experimental results indicate that:

Models trained using proxy losses ( $L_1(y, \hat{y})$ and $MSTE(y, \hat{y})$ ) did not achieve parity with models trained directly on the gain control signal when evaluated against the ground-truth control trajectory.
The spread of errors among models trained with proxy losses was significantly smaller (2.8% for $L_1$ and 13.3% for MSTE) compared to the spread observed when using the direct gain-reduction metric (85.9%).
Waveform-domain metrics assigned similar error values to models that were clearly separated by the direct metric, suggesting that proxy objectives can obscure salient differences in processor behavior.
The direct metric revealed that minimizing proxy losses is not equivalent to minimizing the error in the gain trajectory.

Significance
The paper claims that the inclusion of the control voltage signal in the proposed dataset allows for the direct supervision of the gain reduction behavior, which is the primary quantity of interest in DRC modeling. By enabling evaluation against a ground-truth error defined on the gain trajectory, this work allows researchers to:

Train models using loss functions defined on the true objective, eliminating the need for potentially flawed proxy metrics.
Rigorously evaluate existing proxy metrics to understand their limitations and failure modes.
Investigate specific hardware behaviors (such as the SSL's program-dependent release and curved static response) that generic models may fail to capture.

The authors conclude that while proxy metrics may correlate with the underlying quantity of interest, they are not functionally equivalent surrogates for gain-reduction error, and direct measurement of the control signal offers a more interpretable and accurate basis for model evaluation.