markmuller/TTS_R3_neutral_fidibo_nv

VoxCPM2 Persian — round 3

Continued from markmuller/TTS_POST_trainnig_1_emo / best_checkpoint_of_post_training, 2 epochs (2,917 steps) on 98,061 rows / 270.5 h:

slice	rows	hours	text
gold (TTS_DATA_GOLD)	77,025	141.4	(emotion)[tag] captions, plain text
fidibo50	8,361	49.4	(fidibo) + diacritized text, same-speaker refs
neutral50	2,236	46.6	Gemini long-form (43–150 s), inline per-sentence captions, same-voice refs
nonverbal38k	10,439	33.1	EN + ZH replay with nonverbal tags, zero-shot (CC-BY-NC-4.0 source)

Length-bucketed batches (400 s padded audio per micro-batch, x3 accumulation), LR 1e-5, one B300.

Checkpoints

folder	step	why
step_0002917	2917 (end of epoch 2)	best flow-matching loss on every slice
step_0001000	1000	best stop-head loss — try it if endings are cut short or run on

Each folder is a full trainer checkpoint (weights + optimizer + scheduler), so it can be loaded for inference or resumed.

Per-slice validation (full val sets, identical batches and noise per checkpoint)

diffusion loss / stop loss, lower is better:

ckpt	gold	neutral50	fidibo50	nonverbal38k
init	0.8983 / 0.0137	0.9345 / 0.0023	0.8031 / 0.0038	1.0056 / 0.0255
step 1000	0.8902 / 0.0125	0.9271 / 0.0017	0.8048 / 0.0028	0.9916 / 0.0209
step 2917	0.8837 / 0.0137	0.9235 / 0.0027	0.8035 / 0.0025	0.9901 / 0.0242

run/ holds the generated training config, the training log and the full evaluation JSON.