An Online Tempo Tracker for Automatic Accompaniment Based on Audio-to-audio Alignment and Beat Tracking

We approach a specific scenario in real-time performance following for automatic accompaniment, where a relative tempo value is derived from the deviation between a live target performance and a stored reference, to drive the playback speed of an accompaniment track. We introduce a system which combines an online alignment process with a beat tracker. The former aligns the target performance to the reference without resorting to any symbolic information. The latter utilises the beat positions detected in the accompaniment, reference and target tracks to (1) improve the robustness of the alignment-based tempo model and (2) take over the tempo computation in segments when the alignment error is likely high. While other systems exist that handle structural deviations and mistakes in a performance, the portions of time where the aligner is attempting to find the correct hypothesis can produce erratic tempo values. Our proposed system, publicly available as a Max/MSP external object, addresses this problem.