TTS-AGI/moss-voice-lora-velvet-sage-baritone

LoRA adapters for the Velvet Sage Baritone voice profile

Two LoRA adapters for laion/moss-tts-local-transformer-4.55b-voice-acting-v2, trained on this dataset and on nothing else.

	rank 16	rank 32
alpha	32	64
trainable params	34.36 M	68.71 M
recommended	stage2	stage2

Each rankNN/ directory holds the recommended adapter at its top level, plus both stage1/ and stage2/ so the curriculum can be inspected. RECOMMENDED.json records which stage was chosen and why.

What they were trained on

Nothing but this dataset. Specifically:

• Variants raw and vc_sidon, mixed. Both renderings of each selected candidate are present, so the adapter sees the unconverted take and the converted+restored take of the same performance.
• German and English mixed, in the corpus's natural balance.
• All blocks: emotions, VoiceNet dimensions, edge cases, characters, sports, explicitness.
• Stage 1 -- the better half of every group by the group's own reward ranking (rank < 16 of 32), one epoch.
• Stage 2 -- the winners only (rank == 0 per gid and variant), one epoch, at a lower peak learning rate, warm-started from stage 1.
• 46 groups were held out entirely and never encoded into the training container. The split is by group, because the 32 candidates of a group are the same sentence under the same condition and holding out candidates would leak the line.

Captions are resampled every epoch from the stored per-clip measurements rather than baked in; see PROTOCOL.md.

Usage

from transformers import AutoModel, AutoProcessor
from peft import PeftModel

BASE = "laion/moss-tts-local-transformer-4.55b-voice-acting-v2"
model = AutoModel.from_pretrained(BASE, trust_remote_code=True, dtype="bfloat16")
model = PeftModel.from_pretrained(model, "loras/rank32")

See PROTOCOL.md for hyperparameters, loss curves, GPU-hours and the held-out evaluation after each stage.