Introduction to OpenBPM

OpenBPM is an open-source, high-accuracy BPM (tempo) detection engine written in pure Rust.

Most existing tempo estimation tools pick a single algorithmic method (such as simple peak detection or basic autocorrelation) and hope it works across different musical genres. OpenBPM takes an ensemble approach: it executes six independent signal processing estimators in parallel, clusters their hypotheses, and uses a learned judge router to resolve octave ambiguities.

---

The Two-Stage Architecture

Audio ──► Multi-band filter (low/mid/high)
             │
             ▼
         SuperFlux onset detection (per band)
             │
             ▼
         Weighted merge (kick=2x, snare=1.5x, hat=0.5x)
             │
             ├──► IOI Histogram ──┐
             ├──► Comb Filter ────┤
             ├──► Autocorrelation ┤──► Metrical fusion ──► Octave resolution
             ├──► Low-Band AC ────┤         │
             ├──► Spectral Energy ┘         ▼
             └──► Hopf Oscillator     Judge Router (learned)
                                            │
                                            ▼
                                     Final BPM + Confidence

Stage 1: Pure Signal Processing

Stage 1 uses pure signal processing without learned parameters:

Estimator	What it measures	Why it helps
IOI Histogram	Direct inter-onset interval measurement	Sub-BPM resolution from timing pairs
Comb Filter Bank	Resonance at beat periods	Robust when onsets are missing or noisy
Autocorrelation	Periodicity of onset envelope	Strong on steady, predictable rhythms
Low-Band AC	Kick-only autocorrelation (< 200 Hz)	Immune to triplet hi-hat or syncopated snare confusion
Spectral Energy	FFT of RMS envelope	Completely independent from onset detection
Hopf Oscillator	Nonlinear resonator bank	Robust against syncopation and missing downbeats

Stage 2: Learned Judge Router

A multinomial logistic regression model (32 features, 4 classes) decides whether to keep, halve, double, or triple the Stage 1 BPM estimate:

• Trained on GiantSteps (EDM), Ballroom (dance), and GTZAN (10 genres).
• Only triggers octave adjustment when prediction probability exceeds confidence threshold ($P > 0.65$).
• Weights are embedded directly as compile-time Rust constants: zero external files, zero runtime dependencies.