Note-level Music Transcription by Maximum Likelihood Sampling

Note-level music transcription, which aims to transcribe note events (often represented by pitch, onset and offset times) from music audio, is an important intermediate step towards complete music transcription. In this paper, we present a note-level music transcription system, which is built on a state-of-the-art frame-level multi-pitch estimation (MPE) system. Preliminary note-level transcription achieved by connecting pitch estimates into notes often lead to many spurious notes due to MPE errors. In this paper, we propose to address this problem by randomly sampling notes in the preliminary note-level transcription. Each sample is a subset of all notes and is viewed as a notelevel transcription candidate. We evaluate the likelihood of each candidate using the MPE model, and select the one with the highest likelihood as the final transcription. The likelihood treats notes in a transcription as a whole and favors transcriptions with less spurious notes. Experiments conducted on 110 pieces of J.S. Bach chorales with polyphony from 2 to 4 show that the proposed sampling scheme significantly improves the transcription performance from the preliminary approach. The proposed system also significantly outperforms two other state-of-the-art systems in both frame-level and note-level transcriptions.