PFVN-synth: Synthesized Violin-Piano Ensemble Dataset

A PFVN-synth dataset contains realistic instrumental triplet audios of piano, violin, and their mixture by rendering MIDI files with virtual instruments using musical scores of 45 different pieces by 23 classical composers with a total duration of 7 hours. All MIDI files were collected on the MuseScore website. All tracks are rendered into monaural audio files with the standard CD quality: 44.1kHz, 16-bit. We used commercial virtual instruments to synthesize piano and violin. Specifically, we used ‘Bösendorfer Grand Piano’ in Apple Logic Pro and ‘SWAM Violin V3’ by Audio Modeling.

It is divided into a train set containing 32 pieces with a duration of 5.8 hours, a validation set containing 3 pieces with a duration of 30 minutes, and a test set containing 10 pieces with a duration of 50 minutes so that the composers are not biased to the split sets.