Extracting Sonic Trajectories

Groove in music is more than just rhythmic onsets, it’s a continuous perceptual experience shaped by subtle changes in dynamics, pitch, and timbre. Traditional audio
analysis often focuses on discrete events, missing the evolving character of sound that contributes to immersion and movement. This work proposes a trajectorybased
approach to audio analysis, aimed at extracting sonic trajectories: time series that represent meaningful changes in audio features over time.

To explore this idea, both traditional and modern machine learning-based audio representations are considered, including MFCC, CQT, and latent spaces from neural
models such as Music2Latent and Descript Audio Codec. A library of synthetic audio signals with known modulations was developed, allowing precise ground-truth
comparisons. Various metrics are used to track changes across time in these representations.

An algorithm was built to extract these trajectories by processing representationvectors through smoothing, downsampling, and convexity normalization. A test
bench was created to systematically evaluate how well different representations support trajectory extraction. Results show that some latent spaces are surprisingly
effective in tracking complex modulations, while others struggle with certain types of changes, such as frequency modulation or filtering.

Overall, this work demonstrates the potential of using trajectory-based methods for perceptual audio analysis and provides a framework for testing and comparing
representations in a controlled and replicable way. It also raises important questions about interpretability in learned audio spaces and opens up future directions for
applying this approach to real-world and creative audio systems.