Rethinking the Joint Estimation of Magnitude and Phase for Time-Frequency Domain Neural Vocoders
IEEE Signal Processing Letters · 2026 · 33 · 2770-2774
DOI: 10.1109/lsp.2026.3708025 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Lingling Dai und 4 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Time-frequency (T-F) domain neural. vocoders have shown promising results in synthesizing high-fidelity audio with jointly modeling magnitude and phase representations. Nevertheless, the underlying mechanisms for effective joint magnitude-phase estimation remain insufficiently explored. In this paper, we employ Vocos as a representative T-F vocoder backbone and conduct a thorough study across its architectural variants. Our experimental results demonstrate that the interaction between magnitude and phase streams is an important factor, with the fully shared single-stream variant outperforms its dual-stream counterparts in the baseline setting, highlighting the importance of exploring internal mechanisms and solutions. Based on these observations, we introduce three simple yet effective strategies, targeting respectively the topological space, the source space, and the output space. Specifically, we refine the architectural topology to enhance the information exchange in the topological space, introduce prior knowledge to facilitate the generation process in the source space, and adopt an improved output formulation to promote the more coherent joint estimation. Experimental results show that our approach consistently improves joint magnitude-phase estimation across multiple architectural variants and modeling units, while providing a more general perspective on robust magnitude-phase modeling in T-F neural vocoders.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Dai, L., Li, A., Lei, T., Li, X., & Zheng, C. (2026). Rethinking the Joint Estimation of Magnitude and Phase for Time-Frequency Domain Neural Vocoders. IEEE Signal Processing Letters, 33, 2770-2774. https://doi.org/10.1109/lsp.2026.3708025
Vancouver
Dai L, Li A, Lei T, Li X, Zheng C. Rethinking the Joint Estimation of Magnitude and Phase for Time-Frequency Domain Neural Vocoders. IEEE Signal Processing Letters. 2026;33:2770-2774. doi:10.1109/lsp.2026.3708025
BibTeX
@article{Dai2026,
author = {Dai, Lingling and Li, Andong and Lei, Tong and Li, Xiaodong and Zheng, Chengshi},
title = {Rethinking the Joint Estimation of Magnitude and Phase for Time-Frequency Domain Neural Vocoders},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {2770--2774},
doi = {10.1109/lsp.2026.3708025}
}
EUVIMED