Audio samples from "Combining Multi-Order Attention and Multi-Resolution Discriminator for High-Fidelity Neural Vocoder"
Abstract:
Generative Adversarial Network (GAN-based) vocoders have gained significant attention for generate high-quality audio with fast inference speed. A typical issue with GAN-based models is phase and artifacts in synthesized speech. In this work, we propose a scalable hierarchical architecture incorporating stacked multi-order spatial-channel attention to enhance feature representation learning. This design enables progressive refinement of acoustic features through attention-guided processing. Moreover, we propose a multi-resolution full-band discriminator, which simultaneously captures local spectral details and global waveform coherence with multiple resolutions, effectively mitigating blurring artifacts that commonly occur in conventional discriminators. Additionally, our discriminator explicitly integrates both the magnitude and complex-domain representations by jointly processing the magnitude, real, and imaginary components. The experimental results show that our proposed method outperforms the baseline models in objective and subjective metrics. Audio samples are available online.
1: the forms of printed letters should be beautiful, and that their arrangement on the page should be reasonable and a help to the shapeliness of the letters themselves.
2: The eight courts above enumerated were well supplied with water.
3: that not more than one bottle of wine or one quart of beer could be issued at one time. No account was taken of the amount of liquors admitted in one day.
4: as indefatigable and self-sacrificing, found by personal visitation that the condition of jails throughout the kingdom was.
5: refuted by abundant evidence, and having no foundation whatever in truth.
6: Later on he had devoted himself to the personal investigation of the prisons of the United States.
7: in the face of impediments confessedly discouraging.
8: This contrivance appears to have been copied with improvements from that which had been used in Dublin at a still earlier date,
9: the schoolmaster and the juvenile prisoners being seated round the communion-table, opposite the pulpit.
10: Nor did the methods by which they were perpetrated greatly vary from those in times past.
11: Visited Mr. Fauntleroy. My application for books for him not having been attended, I had no prayer-book to give him.
12: Weedon and Lecasser to twelve and six months respectively in Coldbath Fields.
13: One of the earliest of the big operators in fraudulent finance was Edward Beaumont Smith.
14: This was Delarue's watch, fully identified as such, which Hocker told his brother Delarue had given him the morning of the murder.
15: Chronicles of Newgate, Volume 2. By Arthur Griffiths. Section 18: Newgate notorieties continued, part 3.