Skttttt/Fishspeech-S2-Pro-Itsuki-Edition-V2

Fishspeech S2 Pro "Itsuki Edition" V2 ⭐

V2 of the last model. New and Improved, Trained with lots more data

V2 continues the same personal fan project as Itsuki Edition V1, now with the newer character-only training run and its step-5500 checkpoint. I wanted an Itsuki voice for new Japanese lines, with a reference clip to guide her tone and delivery. ⭐

Built with Fish Audio. This is an unofficial fan project, with no affiliation or endorsement from the creators, performers, or franchise.

Blind listening results: Blind listening tests found V2 to be a dramatic improvement over the base model and V1.

Full merged model

This repository contains the complete BF16 merged model. The trained Itsuki adaptation is already in model.safetensors. No separate LoRA adapter, base-weight download, or merge step is required. Load it as a full model with lora_config=None; do not apply an adapter again.

File	Purpose
model.safetensors	Full merged text-to-semantic weights, 9.92 GB
codec.pth	Original pinned Fish Speech audio codec, 1.87 GB
config.json and tokenizer files	Native dual_ar configuration and tokenizer
infer.py	Wrapper around the pinned Fish Speech inference CLI
references/*.wav, .txt, .npy	Multiple matching reference recordings, transcripts, and codec tokens
LICENSE.md, NOTICE	License and attribution/modification notice

The training corpus, optimizer states, checkpoint files, and standalone adapters are not included.

What changed from V1

V2 ships the newer full merged step-5500 character checkpoint and multiple reference clips. It is a different trained checkpoint from V1.

References

Default: g2_it_10866. All reference recordings are preserved unchanged; each includes its matching transcript and precomputed codec tokens.

g2_it_10866 (default)

9.52 seconds. WAV · Transcript · Tokens

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。

candidate_01

6.08 seconds. WAV · Transcript · Tokens

いやそんな。先ほども言いましたが、あくまで父が所有しているものですから。

candidate_02

7.80 seconds. WAV · Transcript · Tokens

初日に言いましたよね。答えを出さなければいけない。しかし最終日まで時間をくれと。

candidate_03

6.87 seconds. WAV · Transcript · Tokens

もほら二人ともそんなに焦らなくても海は逃げませんよ。のんびり行きましょう。

Choose another included reference with --reference; keep its transcript and tokens together:

python checkpoints/itsuki-v2/infer.py --text "今日は何をして過ごしましょうか。" --reference candidate_01 --output itsuki.wav

Changing the reference can change delivery. The voice is not guaranteed to match the character exactly on every line.

Use the model

Use the compatible Fish Speech fork at commit 166940c301b8b4650b3110397a38bc68727a3afa. This is a native Fish Speech dual_ar model, not a drop-in Transformers AutoModel checkpoint.

Install the fork and a PyTorch build compatible with your hardware using that revision's instructions. The pinned project specifies PyTorch 2.8.0; accelerator support depends on your runtime.

git clone https://github.com/TachibanaKimika/fish-speech.git
cd fish-speech
git checkout 166940c301b8b4650b3110397a38bc68727a3afa
# Follow this revision's platform-specific dependency installation instructions.

Authenticate with hf auth login using an account with access to this private repository, then download the complete model:

hf download Skttttt/Fishspeech-S2-Pro-Itsuki-Edition-V2 --local-dir checkpoints/itsuki-v2

From the installed Fish Speech source directory:

python checkpoints/itsuki-v2/infer.py --text "今日は何をして過ごしましょうか。" --output itsuki.wav

Defaults: included reference, seed 42, BF16, 4,096-token model context, top-p 0.7, top-k 30, temperature 0.7, and at most 1,024 new audio tokens. The helper saves the WAV and generated codec tokens beside it in an output-stem_codes folder. Resource use depends on the text. Downloading the repository does not load or run inference.

Fine-tuning audio

Fine-tuned on approximately 5.27 hours of real recorded Itsuki speech, excluding validation and test audio.

Credits and license

Thanks to Fish Audio, Inori Minase, Negi Haruba, and everyone who brought Itsuki to life in The Quintessential Quintuplets.

The model and derivative weights follow the Fish Audio Research License. See NOTICE for attribution and modification details. Source recordings, performances, character, and franchise remain subject to their respective rights; this project does not claim ownership of them.