In automotive audio systems, musical beats and drum events can control synchronized in-cabin experiences such as ambient lighting and music-driven visual effects. Compared to beat tracking, automatic drum transcription (ADT) offers richer control signals by detecting and classifying drum onsets of multiple drum classes (e.g., kick, snare, and hi-hat), enabling more precise and musically meaningful synchronization. Deploying ADT in vehicles, however, requires low latency, computational efficiency, and robust performance for various input signals. This paper investigates improvements to low-latency ADT suitable for automotive deployment, using the Separate-Tracks-Annotate-Resynthesize Drums (STAR Drums) dataset and a block-based processing strategy that achieves an average detection delay of around 60 ms. We explore three strategies: (1) lightweight architecture modifications inspired by recent advances in image classification, combined with a temporal convolutional network (TCN); (2) re-rendering STAR Drums to increase drum timbre diversity and augmenting the re-synthesized drum stems; and (3) refinement training with pseudo labels obtained from source-separated mixtures. Our results show that data augmentation and increased drum timbre diversity yield modest performance gains, whereas pseudo-label refinement provides the largest effect, with up to 18 % relative improvement in global F-measure. In the real-time eight-class setting, our best model achieves a global F-measure of 0.76 on MDB Drums, competitive with state-of-the-art offline systems, demonstrating that accurate and efficient ADT is feasible for automotive deployment.
