SpeechGenCourse/F5TTS_v1_Base_Emilia_ZH_EN_repro

F5TTS_v1_Base reproduction on Emilia (work in progress)

This is a from-scratch reproduction of F5TTS_v1_Base, which is still training.

• Code, environment and resume instructions: https://github.com/Splend1d/F5-TTS-repro (see repro/README.md)
• W&B: https://wandb.ai/splend1dchan/CFM-TTS/runs/f5tts-main-20260930. Updates 0–5,631 are TensorBoard logs in runs/.

Data note: updates 0–5,500 trained on Emilia ZH+EN (37.8M utterances, ~95k h). From update 5,500 on, the run continues on Emilia EN only (18.1M utterances, 46,643 h, a subset of the same ZH+EN data). The repo name keeps "ZH_EN" for continuity. optim.total_updates=1271061 keeps the original ZH+EN LR schedule and end point.

File	What it is
model_last.pt	Full resumable trainer state (model, EMA, optimizer, scheduler, update). Currently update 40,000 / 1,271,061.
rng_last/rank*.pt	Per-rank RNG state for an exact resume on 4 GPUs. Optional; resume works without it.
config.yaml	Resolved Hydra config of the current (EN-only) run. grad_accumulation_steps depends on the GPU count: NGPU × accum = 16.
vocab.txt	Official F5-TTS pinyin vocab. Use it with pypinyin==0.54.0.
runs/	TensorBoard event files for updates 0–5,631.

Setup: DiT with dim 1024, depth 22, 16 heads, pinyin tokenizer, vocos mel. The batch is 307,200 frames/update and the LR is 7.5e-5 with 20k warmup, then linear decay to update 1,271,061.

To resume, place model_last.pt (and optionally rng_last/) in ckpts/F5TTS_v1_Base_vocos_pinyin_Emilia_EN/ in the repo above and run source repro/state/main.env && bash repro/submit_train.sh. Each new checkpoint replaces model_last.pt; earlier steps are in the commit history (update 5,500 is the 2026-10-02 commit).

Status at update 40,000: past LR warmup, training loss around 0.8–1.1. Not evaluated yet.

License: CC BY-NC 4.0, inherited from Emilia. For non-commercial research only.