F5-TTS v1 Base reproduction

Training from scratch on Emilia ZH+EN (~95k h), target 1.25M updates

Code & setup W&B run Checkpoint (HF, private) Raw data (JSON)
Couldn't load metrics.json.
Update
–
Job
–
Speed
–
Latest checkpoint
–
model_last.pt, resumable

Training loss

Learning rate

Evaluations

UpdateTest setWER ↓SIM ↑UTMOS ↑Seeds
No evaluations yet. The first full eval is planned at 50k updates.

Setup

Model
DiT, dim 1024, depth 22, 16 heads, ff_mult 2, text_dim 512, conv_layers 4 (~336M params)
Data
Emilia ZH+EN @ rev fc71e07: 37.8M utterances, ~95,282 h, pinyin tokenizer, official vocab
Batch
307,200 frames/update (19,200/GPU × NGPU × accum = 16)
Optimizer
AdamW, LR 7.5e-5, 20k warmup, grad clip 1.0, bf16, 11 epochs
Audio
24 kHz, 100-bin vocos mel, hop 256