Arknights-Virtousa

An unofficial Japanese GPT-SoVITS v2ProPlus fine-tune for the voice style of Virtuosa / Arturia from Arknights. The Japanese performance in the source material is credited to Aya Endo (遠藤綾).

This repository is not affiliated with, sponsored by, endorsed by, or approved by Hypergryph, Yostar, the performer, or any other rights holder. Every audio file in this repository is synthetic. It contains no original game audio, transcript archive, or training dataset.

Files

File	Purpose
GPT_weights_v2ProPlus/Arknights-Virtousa-e15.ckpt	GPT semantic model, epoch 15
SoVITS_weights_v2ProPlus/Arknights-Virtousa_e8_s480.pth	SoVITS acoustic model, epoch 8 / step 480
reference/neutral.wav	Synthetic Japanese reference audio for inference
reference/neutral.txt	Exact reference transcript
samples/audiobook.wav	Synthetic audiobook-style demonstration
MODEL_INFO.json	Reproducibility and training metadata
SHA256SUMS	Artifact checksums

The demonstration was generated using only reference/neutral.wav as its prompt audio, so inference does not require an original game recording.

Open the audio file directly

Transcript: 朝の光が差し込むころ、新しい一日が静かに始まりました。

Compatibility

• GPT-SoVITS: v2ProPlus
• Tested upstream commit: d523079fc05d9a8028d6085bffe4a2757c32abb6
• Upstream repository: https://github.com/RVC-Boss/GPT-SoVITS
• Tested runtime: Python 3.10.5, PyTorch 2.4.0 + CUDA 12.1
• Tested GPU: NVIDIA GeForce RTX 4060 Laptop GPU, 8 GB VRAM

The two files here are fine-tuned weights, not a standalone application. The GPT-SoVITS runtime and its official v2ProPlus pretrained dependencies must be installed separately.

Quick Start

Download the repository after accepting its access conditions:

hf auth login
hf download Karaik/Arknights-Virtousa --local-dir Arknights-Virtousa

Open a compatible GPT-SoVITS WebUI and select:

• GPT weight: GPT_weights_v2ProPlus/Arknights-Virtousa-e15.ckpt
• SoVITS weight: SoVITS_weights_v2ProPlus/Arknights-Virtousa_e8_s480.pth
• Reference audio: reference/neutral.wav
• Reference text: the contents of reference/neutral.txt
• Reference language: Japanese
• Target language: Japanese

Validated baseline inference settings are top_k=15, top_p=1.0, temperature=1.0, speed=1.0, and CUDA Graph disabled. CUDA Graph decoding caused occasional repetition on very short Japanese phrases during local testing; normal decoding produced 25/25 non-repeating runs for the phrase used in that test.

For audiobook work, preprocess EPUB ruby into its spoken reading before sending plain text to GPT-SoVITS. This model does not parse EPUB, HTML, or ruby markup.

Training Summary

• Language: Japanese only
• Source: two sets of Japanese in-game Virtuosa voice lines archived by PRTS
• Source material: 76 records, 962.788 seconds
• Training corpus: 117 aligned segments, 910.145 seconds
• Excluded: 22 battle/title or unsuitable segments
• SoVITS: 8 epochs, batch size 2, FP16
• GPT: 15 epochs, batch size 4, mixed FP16
• Whisper usage: word-timing estimation only; it did not replace the attached Japanese transcripts

The reported GPT top-3 training accuracy (0.7742) is a training diagnostic, not an independent quality benchmark.

Limitations

• Emotional range is bounded by the source performances and prompt audio.
• Very short fragments are less reliable than complete Japanese sentences.
• Pronunciation of fictional names and unusual kanji may require kana input.
• Reference choice, punctuation, text segmentation, and random sampling affect pacing and stability.
• The model may reproduce recognizable characteristics of a real performer's recorded performance. It must not be represented as the performer's speech.

Terms and Rights

The weights use the repository-specific conditions in TERMS.md. They are shared for personal, research, and non-commercial use only. Synthetic outputs must be disclosed as synthetic whenever they are shared.

The MIT license of GPT-SoVITS applies to upstream code only. It does not grant rights to Arknights, its characters, source recordings, scripts, or the voice performance. No such underlying rights are granted by this repository. Users are responsible for obtaining any permission required in their jurisdiction.

中文说明

这是非官方的日语 GPT-SoVITS v2ProPlus 微调权重，目标为《明日方舟》阿尔图罗 （塑心）的日语声线，原作日语 CV 为遠藤綾。仓库不包含游戏原始音频、原台词数据 或训练集；附带的参考音频与试听均为模型合成。仅限个人、研究和非商业用途，公开 合成结果时必须明确标注为合成语音，不得冒充角色官方内容或声优本人发言。