Automatic Transcription of Ornamented Irish Traditional Flute Music Using Hidden Markov Models

This paper presents an automatic system for note transcription of Irish traditional flute music containing ornamentation. This is a challenging problem due to the soft nature of onsets and short durations of ornaments. The proposed automatic transcription system is based on hidden Markov models, with separate models being built for notes and for single-note ornaments. Mel-frequency cepstral coefficients are employed to represent the acoustic signal. Different setups of parameters in feature extraction and acoustic modelling are explored. Experimental evaluations are performed on monophonic flute recordings from Grey Larsen’s CD. The performance of the system is evaluated in terms of the transcription of notes as well as detection of onsets. It is demonstrated that the proposed system can achieve a very good note transcription and onset detection performance. Over 28% relative improvement in terms of the F -measure is achieved for onset detection in comparison to conventional onset detection methods based on signal energy and fundamental frequency.