Music Tempo Estimation on Solo Instrumental Performance

Zhanhong He, Roberto Togneri, Xiangyu ZhangarXiv 2025 · DOI · arXiv

Abstract

Automatic Music Transcription (AMT) systems often struggle with accurate tempo estimation in solo instrumental contexts, as most models are trained on multi-instrumental "mixed" audio. This research investigates the impact of instrument-specific training for Temporal Convolutional Networks (TCNs) and the efficacy of various probabilistic post-processing techniques. Our findings indicate that specialized training significantly outperforms generalist models, achieving near-perfect accuracy on solo guitar and doubling performance on complex classical piano pieces.

Introduction

The conversion of audio into symbolic representations like MIDI or sheet music—known as Automatic Music Transcription—requires precise tempo (Beats Per Minute) detection. However, many current systems default to a static 120 BPM because general-purpose tempo estimation models often fail when presented with solo instruments. These "generalist" models are typically trained on pop and rock music where strong rhythmic anchors (e.g., drums and bass) are prevalent. Solo instruments, particularly classical piano with its expressive timing or acoustic guitar with its unique transients, present rhythmic characteristics that these models are not equipped to handle.

Methodology

The study centers on the Temporal Convolutional Network (TCN), an architecture proficient at capturing long-range dependencies in audio signals. The researchers evaluated three primary configurations: pretrained models, instrument-specific training, and specialized post-processing pipelines.

Pipeline of the proposed tempo estimation and post-processing methods, illustrating paths from beat activation to final tempo detection.

As shown in Figure 1, the system explores multiple inference paths to refine raw network outputs. These include Direct Detection of tempo activation, Beat-to-Tempo Inference using Dynamic Bayesian Networks (DBN) to track beats, and Beat Activation Estimation which deduces tempo from raw likelihood signals using Auto-Correlation (ACF) or Comb Filters.

Experimental Setup

Three high-quality solo datasets were utilized for evaluation:

  • ASAP (Aligned Scores and Performances): 44.3 hours of solo classical piano, characterized by expressive, non-linear timing.
  • GMD (Groove MIDI Dataset): 13.2 hours of solo drum performances.
  • GuitarSet: 3 hours of solo acoustic guitar recordings.

To address the limited size of the guitar and drum datasets, data augmentation was performed by varying audio spectrogram frame rates, effectively tripling the training volume.

Results and Analysis

The performance was measured using the Acc1 metric (estimates within ±4% of the ground truth). The results, detailed in Figure 2, highlight a significant "Generalist Penalty" for solo instruments when using models trained on mixed-genre music.

Comparison of tempo estimation accuracy (Acc1 and Acc2) across different models, post-processing methods, and datasets.

Key findings include:

  • Guitar Specialization: The specialized TCN improved guitar tempo accuracy from 61.1% to 99.7%.
  • Piano Complexity: On the ASAP dataset, accuracy doubled from 24.6% to over 50%. The use of a Comb Filter or DBN on beat activations provided a critical boost, helping the model navigate the complex harmonies of classical piano.
  • The Drum Exception: Retraining showed minimal gains for drums (87.9% to 89.3%), likely because generalist models are already heavily exposed to drum-like features in mixed-genre training data.

Conclusion

This research underscores the necessity of instrument-specific specialization for high-fidelity music transcription. While specialized training and post-processing significantly bridge the performance gap, classical piano remains a challenge due to rubato—expressive timing variations that a single global tempo value cannot fully represent. Future work will focus on more sophisticated models capable of tracking these dynamic tempo changes.

Summary updated

BibTeX
@misc{he2025music,
  author = {Zhanhong He and Roberto Togneri and Xiangyu Zhang},
  title = {{Music Tempo Estimation on Solo Instrumental Performance}},
  year = {2025},
  eprint = {2504.18502},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.18502},
}
Figure