Zero-Mean Convolutions for Level-Invariant Singing Voice Detection

State-of-the-art singing voice detectors are based on classifiers trained on annotated examples. As recently shown, such detectors have an important weakness: Since singing voice is correlated with sound level in training data, classifiers learn to become sensitive to input magnitude, and give different predictions for the same signal at different sound levels. Starting from a Convolutional Neural Network (CNN) trained on logarithmic-magnitude mel spectrogram excerpts, we eliminate this dependency by forcing each first-layer convolutional filter to be zero-mean – that is, to have its coefficients sum to zero. In contrast to four other methods – data augmentation, instance normalization, spectral delta features, and per-channel energy normalization (PCEN) – that we evaluated on a largescale public dataset, zero-mean convolutions achieve perfect sound level invariance without any impact on prediction accuracy or computational requirements. We assume that zero-mean convolutions would be useful for other machine listening tasks requiring robustness to level changes.