Deepest-Project/MelNet

Implements a hierarchical multi-tier architecture that generates mel spectrograms through successive refinement, with support for both unconditional synthesis and conditional text-to-speech generation. Built on PyTorch with per-tier training flexibility to manage memory constraints on high-resolution frequency representations. Includes Gaussian mixture model sampling and tensorboard logging, supporting multiple datasets (Blizzard, VoxCeleb2, KSS) via configurable YAML specifications.

210 stars. No commits in the last 6 months.