Research map / Video / Video segmentation and tracking
Skipping the Frame-Level: Event-Based Piano Transcription With Neural Semi-CRFs
Yujia Yan, Frank Cwitkowitz, Zhiyao Duan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove extra adjustable windows (oversampled log-mel) (ablation) | Proposed semi-CRF · MAESTRO v2 note w/ offset | F1: 88.37 -> 88.63 | Improved |
| Remove final convolutional block in score matrix (ablation) | Proposed semi-CRF · MAESTRO v2 note w/ offset | F1: 88.26 -> 88.63 | Improved |
| Remove interval-length scaling of scores (ablation) | Proposed semi-CRF · MAESTRO v2 note w/ offset | F1: 88.51 -> 88.63 | Improved |
| Increase batch size from 12 to 20 (ablation) | Proposed semi-CRF · MAESTRO v2 note w/ offset & velocity | F1: 87.65 -> 87.75 | Improved |
| Split architecture into note and pedal branches sharing only conv blocks (ablation) | Proposed semi-CRF · MAESTRO v2 pedal onset | F1: 78.10 -> 79.98 | Improved |
| Split architecture into note and pedal branches (ablation) | Proposed semi-CRF · MAESTRO v2 pedal onset & offset | F1: 74.71 -> 76.85 | Improved |
| Use event-level semi-CRF instead of frame-level model with post-processing | Proposed vs Kong et al. · MAESTRO v2 note w/ offset | F1: 84.45 -> 88.63 | Improved |
| Use semi-CRF instead of frame-level model | Proposed vs Kong et al. · MAESTRO v2 note onset | F1: 96.77 -> 96.39 | Hurt |
| Use semi-CRF instead of frame-level model | Proposed vs Kong et al. · MAESTRO v2 pedal activation | F1: 94.11 -> 90.73 | Hurt |
| Train on MAESTRO v3 instead of v2 splits | Proposed (batchsize=20) · note w/ offset & velocity | F1: 87.75 -> 87.44 | No clear effect |
| Use Proposed instead of Kong et al. for end-to-end transcription of a 33.3 minute file | Proposed vs Kong et al. · Carl Czerny Grand Sonata Op.145 No.9 | running time (s): 353s -> 95s | Improved |
| Use GPU instead of CPU for forward-backward on 1600 frames | semi-CRF components · runtime | seconds: 7.12 -> 0.27 | Improved |
Closest papers by meaning
- Rubato: Transcribing Piano Music with Timestamps (NeurIPS 2026 · similarity 0.88)
- Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling (ICLR 2025 · similarity 0.86)
- PianoMotion10M: Dataset and Benchmark for Hand Motion Generation in Piano Performance (ICLR 2025 · similarity 0.85)
- Toward Multimodal Sheet Music Recognition and Understanding (NeurIPS 2026 · similarity 0.85)
- Unifying Symbolic Music Arrangement: Track-Aware Reconstruction and Structured Tokenization (NeurIPS 2025 · similarity 0.83)
- Structured Multi-Track Accompaniment Arrangement via Style Prior Modelling (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
