Researchers expose Vocos audio quality gap in neural vocoding

A new paper revisits Vocos, a time-frequency neural vocoder noted for its efficiency but lagging audio quality compared to time-domain models. By quantifying that gap and conducting an ablation study, the authors confirm Vocos excels at magnitude modeling but falls short on phase reconstruction. They adapt its architecture to predict phase differences and find that 1D convolutional layers hinder accurate phase prediction. The study concludes that future research should focus on inductive biases for better time-frequency modeling of speech without compromising input flexibility.