We introduce ChoraleWind, an expressive wind-quartet dataset and reproducible baseline for end-to-end rendering from the Neues Thueringer Choralbuch (NTCB). The dataset comprises 312 four-part chorales and covers the complete path from symbolic score encoding to performance-level rendition and synthesized audio. ChoraleWind includes a rule-based performance model that generates expressive timing, dynamics, and articulation, including metric and structural accents as well as phrase-end gestures from high-quality MEI encoding of the NTCB chorales, combined with controlled humanization, and physical-modeling wind-instrument synthesis that produces isolated voice stems and ensemble mixes. The dataset provides aligned symbolic representations, performance annotations, and multitrack audio, enabling systematic training and evaluation of score-to-audio wind-quartet rendering methods under fully controlled and reproducible conditions. Rather than aiming at state-of-the-art purely data-driven synthesis, ChoraleWind is designed as a transparent baseline and controlled testbed for studying expressive performance generation, timbre modeling, and reproducible evaluation of wind-quartet rendering systems.
pip install uv
uv syncInstructions for rendering the dataset from the provided performance MIDI files are available here:
https://www.audiolabs-erlangen.de/resources/MIR/2026-ChoraleWind/reproduce_dataset