CC-G2PnP: Japanese Grapheme-to-Phoneme and Prosody Model
Conformer-CTC ベースの日本語 G2PnP (Grapheme-to-Phoneme and Prosody) モデル。
Model Details
- Architecture: Conformer Encoder + CTC Head (84M params)
- Training: 300k steps, T4×4 DDP, SDPA enabled
- Checkpoint:
step_00300000.pt(25% of full 1.2M steps training)
Usage
import torch
checkpoint = torch.load("step_00300000.pt", map_location="cpu")