speecht5_tts-pld-pam

microsoft/speecht5_tts finetuned on sapinsapin/pld.

Trained for 1000 steps on 1779 clips (batch 4×8, lr 1e-05, fp32 + gradient checkpointing). Synthesized listen-test samples are in samples/ (speechbrain x-vector speaker conditioning + microsoft/speecht5_hifigan vocoder).

metric	value
eval_loss	0.4330

Trained with finetune_tts.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.