Deep Neural Network for Multi-Pitch Estimation Using Weighted Cross Entropy Loss

Multi-Pitch Estimation, the estimation of multiple overlapping or polyphonic harmonic fundamental frequencies, has a wide range of applications, including automatic music transcription, power systems, and radar signal processing. Multiple fundamental frequencies represent a challenge due to the added complexity of the overlapping signals. This paper presents a Deep Learning approach to estimating multiple fundamental frequencies. The network is trained in a supervised fashion to generate a pseudospectrum representing the fundamental frequencies. Training data is represented by a sparse binary vector the size of the pseudospectrum, indicating the location of fundamental frequencies. A weighted binary cross-entropy loss function is used to correct for class imbalance caused by the sparsity of the signal space relative to the full spectrum. We show comparable performance to existing techniques while requiring fewer operations and samples due to a simpler frequency-domain-only architecture.