Beethoven Symphony Excerpt Dataset (BSED): An Evaluation Dataset for Orchestral Music Transcription

Orchestral music poses significant challenges for automatic music transcription (AMT) due to its dense polyphony, diverse instrument timbres, and varied acoustic conditions. While AMT research for solo instruments and chamber music has advanced considerably, orchestral transcription remains underexplored, primarily because of the lack of high‑quality, time‑aligned datasets. In this work, we make three main contributions. First, we introduce the Beethoven Symphony Excerpt Dataset (BSED), a carefully curated, publicly available benchmark for evaluating orchestral AMT systems. BSED contains 20 short excerpts from Beethoven’s nine symphonies, each provided in five audio versions: four distinct concert recordings and one synthetic rendition, with an approximate total duration of 37 min. For each excerpt, we supply symbolic scores in multiple formats (PDF, MusicXML, Sibelius, MIDI, CSV) alongside high‑quality note‑level annotations obtained through robust, manually verified score–audio alignment, refined with transcription‑based onset features. Second, to facilitate model development, we release the Beethoven Symphony Dataset (BSD), a large‑scale orchestral training set comprising 62 h of public‑domain recordings with time‑aligned annotations derived from digital scores and structurally verified. While less controlled than BSED, BSD offers rich diversity across performances, conductors, and acoustic environments. Third, we establish a baseline for orchestral AMT by training an instrument‑agnostic note‑level transcription model on BSD and evaluating it on both BSED and the independent PHENICX dataset. Together, BSED and BSD address a critical gap in orchestral music information retrieval resources and provide a foundation for advancing AMT research toward more robust and generalizable systems.