Threshold Effect of Neural Network-based Fundamental Frequency Estimators

Recently, neural network (NN)-based fundamental frequency estimators have shown great promise in complex real-world problems such as audio processing, where model-based approaches tend to oversimplify. However, valuable insights can be gained by studying NNs under simpler and more controlled conditions. In this paper, we study the threshold effect, i.e., the rapid decrease in estimation accuracy with decreasing signal-to-noise ratio (SNR), of NN-based fundamental frequency estimators. We use the multiharmonic signal model and simulated data, which allows us to compare the networks with the threshold behavior of the maximum likelihood estimator (MLE) and with the Cramer-Rao lower bound (CRLB) above the threshold SNR. Finally, good practices for the design and training of such networks can be derived from our studies.