You can deploy this model and start earning money today!

xingguangcuican6666/
arona-sovits

Name: arona-sovits
Availability: InStock

Views

No views yet

🎵 Arona v2 — GPT-SoVITS Voice Model

Arona from Blue Archive | 6min training data | Multilingual ZH/JA/EN

Model License

GPT-SoVITS

Hugging Face

简体中文 | English

📦 Model Overview

Property	Value
Character	Arona (Blue Archive)
Version	v2 (6min training data)
Architecture	GPT-SoVITS v2
Languages	Chinese (ZH), Japanese (JA), English (EN)
Sample Rate	32kHz / 44.1kHz
Training Data	~6 minutes of reference audio

Model Files

GPT_weights_v2/
├── speaker01_v2_6min-e5.ckpt      # GPT model (epoch 5)
└── speaker01_v2_6min-e10.ckpt     # GPT model (epoch 10, recommended)

SoVITS_weights_v2/
├── speaker01_v2_6min_e2_s204.pth  # SoVITS (epoch 2)
├── speaker01_v2_6min_e4_s408.pth  # SoVITS (epoch 4)
├── speaker01_v2_6min_e6_s612.pth  # SoVITS (epoch 6)
└── speaker01_v2_6min_e8_s816.pth  # SoVITS (epoch 8, recommended)

Recommended combo: speaker01_v2_6min-e10.ckpt + speaker01_v2_6min_e8_s816.pth

🚀 Quick Start

Option 1: Python (GPT-SoVITS_minimal_inference)

bash

# Install dependencies
pip install onnxruntime-gpu torch

# Clone the minimal inference repo
git clone https://github.com/GPT-SoVITS-Devel/GPT-SoVITS_minimal_inference
cd GPT-SoVITS_minimal_inference

# Download this model
git lfs install
git clone https://huggingface.co/xingguangcuican6666/arona-sovits

# Export to ONNX
python export_onnx.py \
    --gpt_path "arona-sovits/GPT_weights_v2/speaker01_v2_6min-e10.ckpt" \
    --sovits_path "arona-sovits/SoVITS_weights_v2/speaker01_v2_6min_e8_s816.pth" \
    --cnhubert_base_path pretrained_models/chinese-hubert-base \
    --bert_path pretrained_models/chinese-roberta-wwm-ext-large \
    --output_dir "onnx_export/arona_v2" \
    --max_len 1000 \
    --validate

# Optimize to FP16
python onnx_to_fp16.py \
    --input_dir "onnx_export/arona_v2" \
    --output_dir "onnx_export/arona_v2_fp16"

# Run inference
python inference.py \
    --gpt_path "arona-sovits/GPT_weights_v2/speaker01_v2_6min-e10.ckpt" \
    --sovits_path "arona-sovits/SoVITS_weights_v2/speaker01_v2_6min_e8_s816.pth" \
    --ref_audio "arona-sovits/ref.wav" \
    --ref_text "你好，我是阿洛娜。" \
    --text "欢迎来到蔚蓝档案，老师。" \
    --output "output.wav"

Option 2: C++ SDK (Production / Embedded)

For production deployment, use the companion GPT-SoVITS C++ SDK:

bash

git clone https://github.com/GPT-SoVITS-Devel/GPT-SoVITS-cpp
cd GPT-SoVITS-cpp

# Build with CUDA + ONNX Runtime
cmake -B build -S . \
  -DENABLE_CUDA=1 \
  -DONNXRUNTIME_PATH=/path/to/onnxruntime \
  -DCUDA_TOOLKIT_ROOT_DIR=/path/to/cuda \
  -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release

# Create speaker package (.gsppkg)
./build/example/gpt_sovits_cpp_cloud_create_onnx \
    arona \
    ref.wav \
    "你好，我是阿洛娜。" \
    zh \
    arona.gsppkg

# Run edge inference
./build/example/gpt_sovits_cpp_edge_inference_onnx \
    arona.gsppkg \
    "欢迎来到蔚蓝档案，老师。" \
    zh \
    arona \
    output.wav

💡 Why C++ SDK? ~25% faster inference, zero Python overhead, embeddable in apps/games, TensorRT support for even lower latency.

🖥️ Runs on Ancient CPUs (No AVX2)

This SDK has been modified to drop the AVX2 requirement — it runs on CPUs with only SSE4/SSE2, opening up a whole class of legacy hardware as TTS servers.

Real benchmark on an Intel Pentium G645 (2011, 2-core, 2.90GHz):

CPU	Arch	Real-Time?	RTF
Intel Pentium G645	x86_64, only SSE4.2/SSE2	❌ No	~5.0

⚠️ At RTF ≈ 5.0, the G645 takes roughly 5× the audio duration to synthesize — not real-time. It is usable for batch/offline generation, not for live interaction.

Still impressive: even a 2011 entry-level Pentium can run the full multilingual GPT-SoVITS pipeline without AVX2.

🎯 Usage Examples

Reference Audio

Place a reference audio file (ref.wav) in the model directory:

Duration: 3-10 seconds recommended
Format: WAV, 32kHz/44.1kHz, mono
Content: Clean speech in target language

Reference Text (must match audio)

你好，我是阿洛娜。

Inference Text Examples

text

# Chinese
欢迎来到蔚蓝档案，老师。今天也要加油哦！

# Japanese
先生、ブルーアーカイブへようこそ。今日も頑張りましょう！

# English
Welcome to Blue Archive, Sensei. Let's do our best today too!

📊 Model Card Metadata

yaml

license: gpl-3.0
tags:
  - text-to-speech
  - tts
  - gpt-sovits
  - arona
  - blue-archive
  - voice-cloning
  - chinese
  - japanese
  - multilingual
language:
  - zh
  - ja
  - en
library_name: gpt-sovits
model_name: Arona v2
pipeline_tag: text-to-speech

⚖️ License

Component	License
Model weights (`.ckpt`, `.pth`, `.gsppkg`)	GPL-3.0 — see LICENSE_MODEL
Modified C++ SDK code (in this repo)	Apache-2.0 — see LICENSE

⚠️ Important: This model is for non-commercial, research, and personal use only. The character "Arona" belongs to NEXON Games / Yostar. Commercial use requires permission from the IP holders.

🔧 Modified SDK — Upstream & License Declaration

The SDK source code included in this repository is a modified fork of:

GPT-SoVITS-cpp — the production C++ inference engine for GPT-SoVITS.

Item	Detail
Upstream project	GPT-SoVITS-cpp
Upstream license	Apache License 2.0
This fork's license	Apache License 2.0 — see LICENSE
Key modifications	Removed AVX2 instruction-set requirement (works on SSE4/SSE2-only CPUs); custom GPT/SoVITS code & Arona v2 weights integration

In accordance with Apache License 2.0 §4(b), modified files carry prominent notices that they were changed from the upstream source.

🙏 Credits

Base Architecture: GPT-SoVITS by RVC-Boss
Minimal Inference: GPT-SoVITS_minimal_inference
C++ SDK: GPT-SoVITS-cpp
Character: Arona from Blue Archive (NEXON Games / Yostar)
Training Data: Curated from game voice lines (~6 minutes)

📁 Repository Structure

arona-sovits/
├── GPT_weights_v2/           # GPT model checkpoints
│   ├── speaker01_v2_6min-e5.ckpt
│   └── speaker01_v2_6min-e10.ckpt
├── SoVITS_weights_v2/        # SoVITS model checkpoints
│   ├── speaker01_v2_6min_e2_s204.pth
│   ├── speaker01_v2_6min_e4_s408.pth
│   ├── speaker01_v2_6min_e6_s612.pth
│   └── speaker01_v2_6min_e8_s816.pth
├── ref.wav                   # Reference audio (user provided)
├── LICENSE_MODEL             # GPL-3.0 for model weights
└── README.md                 # This file

If this model helps you, drop a ⭐ on the repo. It's free and means a lot! 🤗