Automatic music transcription using accelerated multiplicative update for non-negative spectrogram factorization

Abstract

Music Transcription has been a core field of Music Information Retrieval. Most of the development in MIR is dependent on how efficiently and accurately the notes are extracted. Various methods have been used for Automatic Music Transcription. The most effective ones have been based on Spectral factorization technique for which Non-Negative Matrix Factorization has been profoundly used. Here we use a variant of NMF which is based on Accelerated Multiplicative Update with some predefined templates. The method provides good results on the Disklavier Dataset.