Probabilistic Modelling of Note Events in the Transcription of Monophonic Melodies
This thesis concerns the problem of automatic transcription of music and proposes a method for transcribing monophonic melodies. Recently, computational music content analysis has received considerable attention among researchers due to the rapid growth of music databases. In this area of research, melody transcription plays an important role. Automatic melody transcription can beneﬁt professional musicians and provide interesting applications for consumers, including music retrieval and interactive music applications. The proposed method is based on two probabilistic models: a note event model and a musicological model. The note event model is used to represent note candidates in melodies, and it is based on hidden Markov models. The note event model is trained with an acoustic database of singing sequences, but the training can be done for any melodic instrument and for the desired front-end feature extractors. The musicological model is used to control the transitions between the note candidates by using musical key estimation and by computing the likelihoods of note sequences. The two models are combined to constitute a melody transcription system with a modular architecture. The transcription system is evaluated, and the results are good. Our system transcribes correctly over 90 % of notes, thus halving the amount of errors compared to a simple rounding of pitch estimates to the nearest MIDI note numbers. Particularly, using the note event model signiﬁcantly improves the system performance.