ampixa/limbu-piper-lifwbt

Limbu Piper TTS (LIFWBT)

Work-in-progress Limbu TTS trained from timestamped cleaned Bible audio chunks.

Training shape

• Base: Piper+ multilingual base warmstart
• Data kept: 5,989 chunks / about 18.37 hours
• Speaker layout: one real speaker slot (limbu_bible_voice) plus one unused compatibility slot
• Audio: 22.05 kHz
• Duration predictor: trainable
• WavLM discriminator: disabled for this run

Files

• best/best.ckpt: selected best checkpoint after training finalization
• best/config.json: matching Piper dataset config
• samples/warmstart-base/: fixed evaluation renders before Limbu fine-tuning
• samples/<epoch>/: fixed evaluation renders from different epoch checkpoints
• metadata/: eval manifest, preprocessing stats, speaker estimate, and progress summaries

Best checkpoint selection uses the lowest TensorBoard val_loss by epoch when available; otherwise it falls back to the latest epoch checkpoint.

Current best

• Tag: epoch-0119-step-117360
• Source checkpoint: epoch=119-step=117360.ckpt
• Epoch: 119
• Step: 117360

Uploaded epoch sample checkpoints

• epoch-0229-step-224940
• epoch-0234-step-229830
• epoch-0239-step-234720
• epoch-0244-step-239610
• epoch-0249-step-244500
• epoch-0029-step-29340
• epoch-0039-step-39120
• epoch-0049-step-48900
• epoch-0054-step-53790
• epoch-0059-step-58680
• epoch-0064-step-63570
• epoch-0069-step-68460
• epoch-0074-step-73350
• epoch-0079-step-78240
• epoch-0084-step-83130
• epoch-0089-step-88020
• epoch-0009-step-9780
• epoch-0094-step-92910
• epoch-0099-step-97800
• last-v1