Pitch detection with a neural-net classifier

Pitch detection based on neural-net classifiers is investigated. T o this end, the extent of generalization attainable with neural nets is first examined, and i t is shown t h a t a suitable choice of features is required t o utilize this property. Specifically, invaria n t features should be used whenever possible. For pitch detection, two feature sets, one based on waveform samples and the other based on properties of waveform peaks, are introduced. Experiments with neural classifiers demonstrate t h a t the la t ter feature set which has better invariance properties performs more successfully. I t is found t h a t the best neural-net pitch tracker approaches the level of agreement of human labelers on the same d a t a set , and performs competitively in comparison t o a sophisticated feature-based tracker. An analysis of the errors committed by the neural net (relative to the hand labels used for training) reveals t h a t they are mostly due t o inconsistent hand labeling of ambiguous waveform peaks. Permission to publish this abstract separately is granted. TO APPEAR IN IEEE TRANSACTIONS ON ACOUSTICS, SPEECH & SIGNAL PROCESSING, 1991 .