Automatic Music Transcription dedicated to Chinese Traditional Plucked String Instrument Pipa using Multi-string Probabilistic Latent Component Analysis Models

Abstract

The Probabilistic Latent Component Analysis (PLCA) provides a flexible and highly interpretable framework to model a diversity of music features such as note event specific features (e.g. pitch, duration, amplitude, frequency shifting) and higherlevel knowledge like instrument timbre for Automatic Music Transcription (AMT). In this paper, Multi-String PLCA (MSPLCA) is proposed and allows to model the timbre of individual string characterized by different thickness and materials for polyphonic music transcription of pipa, the head of Chinese traditional plucked string instruments, which is barely studied in the Music Information Retrieval (MIR) community. A dataset, composing 9 famous pieces of Chinese folk music and the notelevel annotation, is created with help of musicians and music experts. As a result, MS-PLCA and its 2 variants adapted to the pipa acoustic features reach AMT performance similar to those found in literature for other instrument transcription. Finally, we illustrate the importance of modeled acoustic features over 2 most common playing techniques, vibrato and tremolo reflecting the periodic pitch and amplitude modulation.