3 months ago
- Teaching machines to follow a drummer's lead involves developing a system where tempo, dynamics, and feel adapt to the drummer, unlike traditional setups where machines dictate timing.
- Simple methods like using kick drum hits as clock pulses fail because real drum patterns include syncopation, fills, and pauses, making them inconsistent for timing.
- A phase-locked loop approach with a beat-aware front end allows a free-running clock to be steered by the drummer's playing, distinguishing between small deviations (feel) and sustained drift (tempo).
- Coasting is essential: the system maintains its last good tempo estimate during confidence wobbles, only stopping on real silence, mimicking how human bandmates respond to drummer variations.
- Scheduling notes ahead based on the clock's forecast, rather than triggering them in response to hits, removes audio latency from the drummer's timing loop, ensuring the band meets the drummer proactively.
- Human timing deviations are correlated and drift with memory, as shown in research by James Holden and Hennig et al., explaining why quantized backing feels dead and a follower's clock must bend gradually.
- Despite advances, human tapping remains superior for tempo tracking due to predictive abilities; algorithms are built because drummers cannot tap while playing and not all jams have extra humans.
- The principles are implemented in DrumMate (drummate.app), an Android app where an e-drum kit triggers a generated band that follows the drummer, with applicability to hardware, software, and plugins.