A Dual-branch Multi-Band Neural Vocoder with Harmonic Discriminator for High-Fidelity Speech Synthesis
Keywords: Dual-branch, multi-band, CondNet, harmonic discriminator
Abstract: Recent developments in vocoders are primarily dominated by GAN-based networks targeting to high-quality waveform generation from mel-spectrogram representations. However, these methods are typically computationally expensive and operate in the time-domain which neglect the time-frequency structures. In this paper, we propose the DMNet, a Dual-branch Multi-band Network to address these limitations. First, a reconstruction network of complex-valued spectrogram called CondNet is used as a condition and thus integrated into the GAN-based branch. Second, we use multi-band processing in the dual-branch: the CondNet produces Fourier spectral coefficients in one sub-band signal and GAN-based branch generates sub-band representations which are subsequently transformed to full-band speech. Finally, to further improve fidelity, we propose a novel harmonic discriminator which utilizes learnable harmonic filters at multiple scales for a better modeling ability in harmonic structures. In our experiments, DMNet validates the effectiveness and achieves superior performance for high quality waveform generation, both on subjective and objective metrics.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3393
Loading