F5-TTS finetune on all formosan data (ithuan, fb ilrdf dict, klokah) without samples only one word, using ipa as input.
Only contains ithuan ami and trv part.
g2p from this repo.
Training Details
learning rate: 0.00001
batch size per gpu: 6400
batch size type: frame
max samples: 64
grad accumulation steps: 1
max grad norm: 1
epochs: 210 (1704780 steps, current 1081600), after 1081600 loss rise
