Noise-Dependent Music Dynamic Range Control

This thesis project explores a psychoacoustic dynamic range control method for music playback in noisy environments. Instead of simply raising volume, the system estimates how environmental noise masks music, analyzes the music with a constant-Q Gabor transform, and applies a time-frequency gain map to selectively restore masked details.

Background

Everyday playback environments introduce nonstationary background noise, such as cafe noise, car cabin noise, and speech-like interference. These sounds can mask low-level musical details and make a simple volume boost feel unnatural. The goal of this project was to restore masked details while preserving the original dynamics and tonal balance of the music.

Gabor Analysis

Time-domain Gabor basis function at 39 Hz with 1/12 octave resolution.
A low-frequency Gabor atom. Each atom is localized in time and frequency, allowing musical content to be modified in small perceptual regions.

The music signal is represented as a set of Gabor atoms. Each atom captures energy around a specific time and frequency region, so the algorithm can adjust masked musical details without applying the same gain change to the entire signal.

Sparse Gabor Coefficients

Sparse Gabor sampling lattice showing a non-uniform time-frequency coefficient structure.
Sparse Gabor coefficient structure from the thesis. Lower frequencies use longer atoms, while higher frequencies are sampled more densely in time.

The algorithm uses a constant-Q Gabor representation, where each coefficient corresponds to a localized time-frequency atom. This sparse structure makes it possible to work directly on musical components that are likely to be masked, instead of applying broad static EQ or compression to the entire signal.

Masking Threshold

Masking threshold pipeline showing Bark power, spreading function, tonality offset, p-sum masking threshold, and final threshold with absolute threshold of hearing.
Masking threshold estimation in the Bark domain. The final threshold estimates which parts of the music are likely to be hidden by environmental noise.

A masking threshold describes the level a sound must exceed to remain audible in the presence of another sound. In this project, environmental noise is analyzed in Bark bands, spread across neighboring critical bands, adjusted by tonality, and combined with the absolute threshold of hearing. The resulting curve becomes a perceptual boundary: music components below or near the threshold are candidates for selective enhancement.

Processing Pipeline

Gabor-domain processing pipeline showing original music PSD, noise masking threshold, smoothed gain map, and processed PSD.
Processing example with an Ocean Piano excerpt in cafe noise: original Gabor PSD, masking PSD, smoothed gain map, and mapped Gabor PSD after enhancement.

The processing pipeline was built in MATLAB with C++ MEX bindings to the Gaborator library. Environmental noise is converted into a Bark-domain masking threshold, while the music signal is processed in a higher-resolution Gabor domain so that gain can be applied only where it is perceptually useful.

Subjective Evaluation

Boxplots showing overall listener preference grouped by vocal, classical, jazz, and piano music.
Overall listener preference grouped by music type. Classical music showed the clearest preference for the proposed processing, while vocal material was closer to neutral.

The method was evaluated with objective spectral distortion metrics and a double-blind listening test with 17 participants. The results were content-dependent: classical music benefited the most, jazz and piano showed moderate preference, and vocal material was closer to neutral. Overall, the processed signal was preferred more often than a simple volume-boost baseline in challenging noisy conditions.