trail mix

• beat salad estimates global BPM, beat positions with downbeat inference, and local tempo segments. Uses autocorrelation with harmonic scoring and a gentle octave prior.
• key lime estimates global and local major or minor keys using dual chroma extraction (Goertzel + Constant-Q Transform), multiple key profiles (Krumhansl-Kessler, Temperley, EDMA, and learned), median aggregation, and confidence-gated segmentation.
• sampler platter generates compact min/max/RMS waveform columns.

The trailmix crate combines the three analyzers behind one PCM-in/results-out API. File decoding is kept in the optional trailmix-codecs crate.

Supporting crates:

• trailmix-cli: command-line analysis of individual files.
• trailmix-bench: machine-readable benchmark reports (MIREX-weighted key score, beat F1, Serato agreement, timing).
• trailmix-manifest: shared corpus annotation schema.
• trailmix-datasets: imports GiantSteps reference annotations.
• test-kitchen: loopback browser UI for playback, annotation, and on-demand analysis.

trail mix requires Rust 1.85 or newer. To try the CLI from the repository root:

cargo run -p trailmix-cli -- path/to/audio.flac

The CLI supports MP3, FLAC, AIFF, WAV, AAC-in-MP4, and ALAC-in-MP4 and prints versioned analysis results as JSON. CI tests Linux, macOS, and Windows.

• Pure Rust, zero ML dependencies: all analysis is deterministic DSP.
• Small binary: all three analyzers fit in under 3 MB.
• Portable: runs on ARM, x86, and WASM without platform-specific inference machinery.
• Interpretable: intermediate features (chroma vectors, onset envelopes, confidence scores) are exposed, not hidden in a black box.

trail mix uses heuristic DSP, not trained models. Accuracy is below ML-based systems (CNN/transformer SOTA reaches ~73-78% key accuracy on GiantSteps). The tradeoff is explicit: lower accuracy in exchange for zero deployment complexity, no model downloads, and full algorithmic transparency.

See architecture for algorithm descriptions and design rationale.