KRAISLER Dataset KRAISLER (KAIST Realistic Annotated Instrument-Specific Live Ensemble Recordings) dataset is a multi-track dataset of piano and violin duet recordings for music information retrieval (MIR) research. The dataset contains 20 excerpts from Western classical pieces spanning a wide range of historical periods and genres, performed by a professional violinist and pianist. Each excerpt was recorded live in two acoustically isolated rooms, yielding clean, bleed-free piano and violin stems together with their mixture. Each excerpt is paired with directly captured piano performance MIDI, MusicXML and PDF scores, piano score-to-MIDI alignment, violin note annotations, and global beat annotations. The audio is provided under three acoustic conditions: dry, studio, and hall. Additional information about each piece — composer, title, key, year of composition, and duration — can be found in "metadata.csv". Dataset Structure The dataset is organized into folders by data type. Every excerpt shares an identical two-digit track ID (`01`–`20`) as its file-name prefix, so files that belong to the same piece can be matched across folders. Instrument-specific files use the `PF` (piano) and `VN` (violin) suffixes. Folder Contents Format Files per track performance_wav/ Piano, violin, and mixture audio, each in dry / studio / hall conditions .wav 9 performance_midi/ Piano performance MIDI captured by Disklavier piano .mid 1 score_musicxml/ Full score, piano part, violin part .musicxml 3 score_pdf/ Full score rendered from MusicXML .pdf 1 alignment_match/ Piano score-to-MIDI note alignment .match 1 annotation_csv/ Violin note annotations and global beat annotations .csv 2 Each format represents the following information: - performance_wav: Audio recordings in 44.1 kHz 16-bit WAV format. The piano is recorded in stereo, the violin in mono, and the mixture in mono. Three acoustic versions are provided per instrument following the naming pattern `NN_{PF,VN,mix}_{dry,studio,hall}.wav`. - performance_midi: Piano performance data captured directly from a Yamaha Disklavier, including note onsets, offsets, velocities, and pedal usage (`NN_PF.mid`). - score_musicxml: Symbolic scores in MusicXML format. Each track provides a full score (`NN.musicxml`), a piano part (`NN_PF.musicxml`), and a violin part (`NN_VN.musicxml`). Repeated sections in the original score are fully expanded. - score_pdf: The full score rendered from MusicXML as a PDF image (`NN.pdf`). - alignment_match: Note-wise score-to-MIDI alignment for the piano part in `.match` format (`NN_PF.match`), loadable via the [partitura](https://github.com/CPJKU/partitura) library. - annotation_csv: Violin note annotations (`NN_notes_VN.csv`) and global beat annotations (`NN_beats.csv`) in comma-separated value (CSV) format. Audio Recording The recordings were made in two adjacent, acoustically isolated rooms separated by a soundproof window that enabled visual communication between the performers, resulting in a bleed-free recording setup. Both performers wore wired headphones for real-time monitoring of each other to ensure precise synchronization. The violin was captured with a Shure SM57 dynamic microphone and the piano with a matched stereo pair of large-diaphragm condenser microphones, recorded at 44.1 kHz / 16-bit through a Focusrite Clarett+ 8Pre interface using Logic Pro. Each excerpt was recorded two to three times, and the best takes were selected. To simulate realistic acoustic environments beyond the dry recordings, reverb was applied independently to each instrument stem using AudioEase Altiverb 8, yielding three versions: - dry — the original close-miked recording, - studio — 'Paramount Stage' impulse response, - hall — 'London Wigmore Hall' impulse response. Equalization, compression, and limiting were tailored per piece rather than applied uniformly, under the initial guidance of a professional sound engineer. The final mixture was obtained by summing the processed piano and violin stems at a predefined signal-to-noise ratio of 0 dB (piano as signal, violin as noise), with a limiter applied to prevent clipping. Note Annotations Piano (match file)A note-wise score-to-MIDI alignment was performed for the piano part using the finalized piano MusicXML score and the performance MIDI. Alignment was generated with the [parangonar](https://github.com/sildater/parangonar) library using hierarchical piano-roll DTW, producing `.match` files that annotate each note with its MusicXML and MIDI note IDs, channel, pitch, and onset time. Violin (csv file) Violin note annotations were extracted automatically via a score-to-audio alignment procedure (NoteEM), which simultaneously trains a transcriber and warps the reference violin score to the audio using an Expectation-Maximization scheme with pitch-shift augmentation. The automatic annotations were subsequently reviewed and manually refined by a musically trained annotator. Each row contains one note: Column Description onset Note onset time in seconds offset Note offset time in seconds midi MIDI note number velocity Fixed at 127 for all notes Because the violin does not provide a direct MIDI capture mechanism, offset estimates are inherently less reliable than onset and pitch values, and velocity is not available. Beat Annotations A single global beat annotation is shared across both instruments, rather than separate beat tracks per instrument. Initial beat positions were extracted from the piano `.match` alignment and then manually refined by annotators with formal musical training to reflect the expressive timing of the duet. A set of guidelines was followed to handle ensemble-specific cases (e.g., taking the midpoint of two onsets after a visual cue, following the onset of the rhythm part when the two instruments play melody and rhythm separately, and ignoring grace notes and arpeggios that precede the beat). Each row contains one beat: Column Description beat_time Beat position in seconds beat_type `db` for a downbeat, `b` for a regular beat time_signature Time signature (e.g., `3/4`); populated only when it changes, otherwise blank piano Boolean flag — whether the piano is active at this beat violin Boolean flag — whether the violin is active at this beat The instrument flags mark solo passages explicitly; an instrument is labeled inactive (`False`) only when it is clearly silent. License This dataset was created by the KAIST Music and Audio Computing Lab. KRAISLER is released under the Creative Commons Attribution 4.0 International license. It is provided primarily for research purposes and may not be used for commercial purposes. When sharing results based on KRAISLER, any act that defames the performers is strictly prohibited. Contact We would love to hear how you are using KRAISLER. If you find the dataset useful, please reach out to us at kaist.mac@gmail.com and tell us about your work.