Real-Time Monophonic Dual-Pitch Extraction Model

Pitch estimation is a fundamental aspect of audio signal processing with applications in music information retrieval (MIR), speech analysis and more. This paper proposes a new model for real-time dual-pitch estimation to tackle the problem of estimating two pitches at the same time in duet songs. The model is designed for real-time processing, frame by frame and with streamlined components for efficiency. We also propose data augmentation methods for simulating duet singing scenarios. Our model is robust in single and dual-pitch scenarios through experiments with various datasets and metrics. Our results contribute to the pitch estimation techniques and provide a practical solution for real-time audio processing in MIR, speech analysis and beyond.