Skttttt/Fishspeech-S2-Pro-Itsuki-Edition-V1

Fishspeech S2 Pro "Itsuki Edition" ⭐

I made this because I love Itsuki. Simple as that.

Well I like Itsuki, No reason at allll

I wanted to get some motivation from my favorite character so I put together an Itsuki only fine tune of Fish Audio S2 Pro. And I'm happy enough with the result to share it and if you like Itsuki too I hope you enjoy trying it out. ⭐

You can give it new Japanese text and one of the reference clips below to generate speech (or find a new one on your own) . It won't sound exactly like her every time, the reference and the line you're asking her to say both make a difference.

Built with Fish Audio. This is a fan project. It isn't an official release or endorsed by anyone involved with The Quintessential Quintuplets.

What is included

This is the full merged model, with model.pth, codec.pth, configuration, and tokenizer files included. The Itsuki fine-tune is already merged into the weights; no separate adapter or merge step is needed to run it.

File	Purpose
model.pth	Merged text-to-semantic model, approximately 9.95 GB
codec.pth	Fish Speech audio codec, approximately 1.87 GB
config.json and tokenizer files	Model configuration and text tokenizer
references/	Three audio references, each paired with its transcript
infer_itsuki.py	Command-line helper for selecting a reference and running inference

The training corpus itself is not included.

Reference clips

I included three clips so you can try them and see which one you prefer. The names come from the content in each line. Switching the reference can change the pace and expression quite a bit.

Starlit Navigation (starlit-navigation)

A line about travelling by the stars and remembering something Fuutaro taught her. 9.52 seconds.

海の上では星を目印に旅をしていたのでしょう。去年の夏島に行ったとき上杉くんが教えてくれたことです。

Town Stroll (town-stroll)

A line about enjoying the town and the feeling of travelling abroad. 8.03 seconds.

本当に面白いところでしたね。歩きながら街並みを見ているだけでも海外旅行の雰囲気を味わえました。

Graceful Sea Turtle (graceful-sea-turtle)

Itsuki talking about sea turtles. What can I say. The helper uses this one by default.

ウミガメの泳ぎはテレビやネットで見たことがありますがとても優雅なんですよ まるで空を飛んでいるかのようでした

Always use the .txt belonging to the selected .wav.

Use the model

This checkpoint uses Fish Speech's dual_ar checkpoint layout. Run it with the compatible Fish Speech source; it is not a drop-in Hugging Face Transformers AutoModel checkpoint.

The local release was built and run with Fish Speech commit e5e292632cb11e7a27b2b7487f58f612bc101e13. Follow that revision's installation instructions for your platform. GPU inference requires a compatible PyTorch installation.

Download this repository into checkpoints/itsuki-nakano inside the Fish Speech source directory:

hf download Skttttt/Fishspeech-S2-Pro-Itsuki-Edition-V1 --local-dir checkpoints/itsuki-nakano

From the Fish Speech directory, using its Python environment:

python checkpoints/itsuki-nakano/infer_itsuki.py \
  --text "今日は二次関数について勉強しましょう。まず、公式を確認してください。" \
  --reference graceful-sea-turtle \
  --output itsuki.wav

To try a different reference, set --reference starlit-navigation or --reference town-stroll. The helper automatically reads the matching transcript. It uses a 4,096-token sequence limit and seed 42 by default.

On Windows, the equivalent single-line command is:

.\.venv\Scripts\python.exe checkpoints/itsuki-nakano/infer_itsuki.py --text "今日は二次関数について勉強しましょう。" --reference starlit-navigation --output itsuki.wav

Training details

Setting	Value
Base model	Fish Audio S2 Pro
Training speaker	Itsuki only
Training clips	2,389
Training audio	Approximately 3.13 hours
Adaptation	Fast autoregressive transformer only
LoRA rank / alpha	4 / 2
LoRA dropout	0.01
Target modules	fast_attention, fast_mlp, fast_output
Trainable parameters	Approximately 0.86 million
Shipped adapter	Step 1200

The slow transformer and token embeddings remained frozen. These details describe the original shipped training run; the larger collection of 5,038 clips supplied later was not used to train this checkpoint.

Credits and license

Thanks to Fish Audio for S2 Pro and Fish Speech, and to Inori Minase, Negi Haruba, and everyone who Worked on Itsuki in The Quintessential Quintuplets. She is a big part of why I wanted to make this in the first place.

The model and its derivative weights are distributed under the Fish Audio Research License. Research and non-commercial use are permitted under its terms. Commercial use requires a separate license from Fish Audio. See NOTICE for attribution and the modification notice.

The reference recordings and the character remain subject to their respective rights. The model license does not grant separate rights to the source performances, character, or franchise, and this fan project does not claim ownership of them.

⭐⭐