Reproduce “Comparing Conventional Pitch Detection Algorithms with a Neural Network Approach” on OpenResearch

About this paper

Anja Kroon

This research from McGill University systematically compares two traditional pitch detection algorithms, pYIN and YAAPT, with a neural network-based method, CREPE, using a laryngeal signal-derived ground truth dataset. The study found that traditional algorithms, particularly YAAPT, generally outperformed or matched the neural network approach in terms of voicing decision accuracy and fine pitch error, with pYIN showing the most consistent performance.

Key insights

• Well-designed conventional algorithms like YAAPT and pYIN can perform comparably to or better than neural network approaches for pitch detection in speech.
• Different algorithms excel in specific aspects: YAAPT for accurate voicing decisions, pYIN for consistent performance with uncertainty estimates, and CREPE (despite training on music) shows comparable accuracy but with higher computational needs.
• The domain of training data significantly impacts neural network performance, as CREPE's training on music rather than speech likely limited its performance on speech signals.

Summary by alphaXiv.