souba67/voxcpm-sunda-lora-full

VoxCPM1.5 Sundanese LoRA (full SLR44)

LoRA adapter (r=64, alpha=64, DiT + LM layers) for openbmb/VoxCPM1.5, fine-tuned for Sundanese text-to-speech on the full OpenSLR SLR44 corpus: su_id_female + su_id_male, 4,213 utterances (~5.4 h, 41 speakers, 48 kHz studio recordings, resampled to 44.1 kHz), with same-speaker reference-audio pairing on 40% of rows.

Training

• 8,000 steps (7.7 epochs), batch 4, lr 1e-4, bf16, single Kaggle T4.
• Final val loss 0.628 (loss/diff) at step 8,000.
• Train/val split 4,173 / 40. Full config in train_config.yaml.
• Trained with the official VoxCPM train_voxcpm_finetune.py LoRA recipe.
• Kaggle run: muhammadshobir/phase-f5-voxcpm-full-slr44 version 3.

Measured intelligibility (ASR-WER-as-judge)

Judge: souba67/whisper-tiny-sundanese (itself ~10.6% WER on SLR44-held-out speech). Frozen 11-sentence eval set (10 short + 1 long-form probe), lowercased, punctuation stripped, word-level Levenshtein. Lower is better; this measures intelligibility to one ASR judge, not naturalness.

system	WER
VoxCPM1.5 base, zero-shot Sundanese	85.47%
this adapter	47.86%

For context, on the same eval set an MMS-VITS full fine-tune scored 50.68% (10 short sentences only) and a CSM-1B LoRA 64.96%. The long-form probe did not collapse with this adapter (17/44 word errors vs 30/44 for the base).

No human MOS evaluation yet; treat quality claims accordingly.

Usage

Load with the VoxCPM repo's inference script, passing this adapter as --lora_ckpt on top of openbmb/VoxCPM1.5 weights.