Coupled tensor factorization models for polyphonic music transcription

Abstract

Generalized Coupled Tensor Factorization (GCTF) is a recently proposed algorithmic framework for simultaneously estimating tensor factorization models where several tensors can share a set of latent factors. This paper presents two models in this framework for transcribing polyphonic piano pieces. The first model is based on Non-negative Matrix Factorization where the coupling provides the spectral information to the model. As an extension to the first model, the second model incorporates temporal and harmonic information by taking a rough, incomplete transciption of the piece as input. Incorporating harmonic knowledge improves the transcription quality as the the experimental results show that we get around 23 % F-measure improvement on real piano data.