Monophonic music transcription using neural semi-Markov conditional random field with monophonic path inference

This paper addresses automatic transcription of monophonic music, a fundamental but still important task. A previous method using a neural semi-Markov conditional random field (NS-CRF) achieved state-of-the-art performance in polyphonic piano transcription. Its decoder has two advantages: it models notes as closed intervals and finds the maximum likelihood solution without approximation. It introduces an inductive bias by deriving note pseudo-probabilities per pitch from the inner product of onset and offset variables, simplifying interval evaluation to a relation between two time points. However, applying this to monophonic music introduces computational redundancy due to its stricter constraints. We propose modifications to the path inference procedures of the NS-CRF, specifically adapting the forward–backward algorithm for computing the partition function and the Viterbi algorithm for maximum likelihood decoding, to suit monophonic constraints. These modifications aim to restrict the search space of trainable parameters, thereby improving training efficiency. In addition, we strive to reduce computational cost by jointly computing note pseudo-probabilities. Experiments on monophonic music datasets demonstrate that our decoder improves training efficiency, suggesting its effectiveness for monophonic music transcription. [Work supported by JSPS grants 24H00715 and 25H01148.]