Search papers, labs, and topics across Lattice.
National Institute of Information and Communications Technology
2
0
3
Speaker-disentangled tokenization boosts syllabic segment accuracy and enhances language model performance, revealing a critical link between speaker identity and linguistic content.
A single architecture now unlocks both fast GAN-based and diffusion-based neural vocoders, achieving state-of-the-art speed and quality while drastically reducing training time.
