Overview
trail mix is a local audio-analysis library. It has three independent analyzers for tempo and beats, key, and waveform. All three accept decoded mono PCM. The combined API runs them together and returns an aggregate result. Usage documents these interfaces, and Architecture describes their execution and result types.
Tempo and beat tracking use deterministic DSP. A 24-band spectral-flux onset function feeds autocorrelation-based tempo estimation and dynamic-programming beat placement. Across the GiantSteps full-track corpus and the FSL10K short-loop corpus, global BPM falls within 4 percent of the reference for about 40 to 70 percent of files. Accepting half- and double-time equivalents raises the range to about 60 to 80 percent.
Waveform analysis groups the PCM into display columns and stores the minimum, maximum, and RMS energy for each column. These values retain the information needed to draw a detailed overview without storing the full PCM.
Key estimation combines DSP with a compact learned component. DSP computes tuned twelve-bin pitch-class profiles, scores the 24 major and minor keys, and identifies local key segments. For global key, profile scoring reached a MIREX key score of about 61 on a sealed holdout. A compact transformer reranker using the same pitch-class sequence and DSP scores raises the MIREX key score to about 74, close to larger learned baselines on the same corpus. The transformer is used only for global key selection. Its weights and forward pass are embedded in the Rust binary, so the release build does not require a separate inference framework.