This paper presents an optimization and training strategy analysis for improving singing voice separation using a deeper STFT-based U-Net architecture. The proposed framework utilizes Short-Time Fourier Transform (STFT) spectrogram representations and investigates the impact of different normalization methods, loss functions, optimizers, and activation functions on vocal separation performance using the MUSDB18 benchmark dataset. The study evaluates multiple experimental configurations involving Adam and AdamW optimizers, MSE, MAE, Huber, and combined IRM+IBM loss functions, along with ReLU and LeakyReLU activations. Frequency-axis Min-Max normalization combined with AdamW, LeakyReLU, and the proposed IRM+IBM loss achieved the best performance with an SDR of 7.6 dB, SIR of 26.1 dB, and SAR of 7.7 dB. The results demonstrate that optimization strategies and training configurations significantly influence vocal reconstruction quality, interference suppression, and overall singing voice separation performance in spectrogram-based deep learning systems.
Ajay Mahdoriya Satyam Kanojiya (Thu,) studied this question.