speecht5_vc-pld

microsoft/speecht5_vc finetuned on sapinsapin/pld.

Any-to-any voice conversion across 10 Philippine languages (bcl, ceb, eng, fil, hil, ilo, pag, pam, tsg, war). Trained for 1500 steps on 3088 same-sentence cross-speaker pairs mined from PLD's shared prompt lists (batch 2×16, lr 1e-05, fp32 + gradient checkpointing), conditioned on the target clip's speechbrain x-vector. samples/ holds held-out conversions: *_src → *_converted vs *_target_ref.

metric	value
eval_loss	0.4588

Trained with finetune_s2s.py from the halohalo pipeline; the dataset adapter normalizes each corpus to (audio@16k, text, speaker_id) so corpora are swappable with a --dataset flag.