Noise-robust Pitch Detection Based on Super-Resolution Harmonics

In order to improve the performance of pitch detection algorithms in a noisy scenario, we propose a noise-robust pitch detection model based on super-resolution harmonics. The model consists of denoise, integration, and super-resolution modules. The Denoisy module highlights the harmonic structure of speech in the frequency domain to reduce noise interference. The enhanced spectrum is multiplied by an integral matrix to obtain a pitch-harmonic integration spectrum. The super-resolution module adjusts the low-resolution pitch-harmonic integration spectrum and outputs fine results. In addition, dynamic programming is employed for pitch tracking to eliminate those singular spots in the preliminary results. The experimental results show that the proposed model significantly outperforms the referenced approaches, including SWIPE, CREPE, HarmoF0, and HGCN+. The results also show that the model can work well even under low SNR.