Reproduce “Improving Neural Pitch Estimation with SWIPE Kernels” on OpenResearch

About this paper

David Marttila, Joshua D. Reiss

Research from the Centre for Digital Music at Queen Mary University of London investigates using Sawtooth-Inspired Pitch Estimation (SWIPE) kernels as a specialized frontend for neural pitch estimators. Their approach demonstrates superior or competitive accuracy compared to state-of-the-art models, achieving this with significantly fewer parameters and enhanced robustness to noise.

Key insights

• The Sawtooth-Inspired Pitch Estimation (SWIPE) algorithm, when meticulously implemented, achieves higher accuracy than commonly acknowledged in neural pitch estimation literature, often outperforming state-of-the-art self-supervised neural models.
• Integrating task-specific digital signal processing (DSP) features, such as SWIPE kernels, as a frontend for neural networks allows for highly parameter-efficient models that maintain competitive or superior pitch estimation accuracy.
• Combining classical DSP methods with deep learning can lead to models with improved robustness to noisy environments and offer flexible control over the latency-accuracy trade-off without requiring model retraining.

Summary by alphaXiv.