etts-lora-repo

Emotional VITS LoRA Adapters

This repository contains LoRA adapter checkpoints for a VITS-based TTS model (base: kakao-enterprise/vits-ljs). Each adapter captures a different emotional style and is stored alongside its adapter configuration.

Included adapters

• vits_lora_neutral/ — neutral speech adapter (contains adapter_model.safetensors and adapter_config.json)
• vits_lora_happy/ — happy speech adapter
• vits_lora_sad/ — sad speech adapter
• vits_lora_angry/ — angry speech adapter
• vits_lora_surprise/ — surprise speech adapter

Files of interest

• *.safetensors — LoRA adapter weights (tracked with Git LFS; see .gitattributes).
• adapter_config.json — PEFT-compatible adapter configuration (generated by PEFT).

Quick Usage

1. Install dependencies (example):

pip install transformers accelerate peft safetensors soundfile phonemizer

1. Basic loading pattern (illustrative):

from transformers import AutoModel, AutoProcessor
from peft import PeftModel, PeftConfig

# Load base model and processor (replace with your exact VITS classes)
base = AutoModel.from_pretrained("kakao-enterprise/vits-ljs")
# Wrap with PEFT adapter
peft_model = PeftModel.from_pretrained(base, "./vits_lora_happy")

# Use the model as you normally would for inference
# (model-specific generation/inference code not included here)

Notes

• This repository uses Git LFS for large model files. The file .gitattributes lists extensions tracked by LFS (e.g. *.safetensors). If you clone the repo and don't have LFS installed, install it or download the weights manually from the release/storage.
• The adapter configs were produced by PEFT (see peft_version in adapter_config.json). They are compatible with peft 0.19.x and later but should work with newer releases as well.

Uploading to Hugging Face Hub

• If you want to publish an adapter as a model on the Hugging Face Hub, follow these steps:

1. Create a new model repository on Hugging Face (e.g. username/vits-lora-happy).
2. Install huggingface_hub and login: huggingface-cli login.
3. From this repo folder, push the adapter directory (for example vits_lora_happy) using the hub API or git:

cd vits_lora_happy
git init
git remote add origin https://huggingface.co/username/vits-lora-happy
git add .
git commit -m "Add vits LoRA happy adapter"
git lfs install --local
git push origin main

Or use huggingface_hub python API to upload individual files if you prefer.

License & Attribution

• This repository contains adapter weights derived from kakao-enterprise/vits-ljs. Respect the original model license when redistributing or publishing derivative weights.

Contact / Next Steps

• If you'd like, I can add example inference scripts, a minimal Dockerfile, or a notebook demonstrating how to apply these adapters to synthesize audio. Tell me which you'd prefer.