Iroha-TTS-V1

Give every line a little more character.

Expressive Japanese TTS for character dialogue. Clone a voice, choose an emotion, adjust its intensity, and explore different takes on the same line. Built with Fish Audio. Based on Fish Audio S2-Pro.

Features · Quick start · Parameters · Agent guide · 中文

Features

01 · Direct the emotion Give one character joy, anger, sadness or surprise with a simple tag. [happy] [angry] [sad] [surprised]	02 · Dial in the intensity Choose one of four presets, or pass a float within the supported range to explore the space between them. weakest → weak → default → strong intensity=-0.015
03 · Turn the mood mid-line Move from delight to sadness within one generation, without stitching separate audio clips. [happy]やっと会えた。[sad]でも、もうお別れだね。	04 · Your voice, in character Clone a voice from a short reference recording and its transcript, then explore different expressions. Reference audio + transcript
05 · More than spoken words Place laughter, sighs or crying where the scene calls for them. [laugh] [sigh] [cry]	06 · Tell it in a whisper Use a whisper tag for quiet dialogue and close-up moments. [whisper]ここだけの話だよ。
07 · Give names their reading Add kana to names, rare words and heteronyms to specify how they should be pronounced. 水神[すいじん] / 水神[みずかみ]

Intensity control is experimental: continuous coefficients do not imply a linear perceptual response. Crying, whispering and voice matching vary with the line and reference. The illustrations show expressions; the reference audio determines the voice. Hear more scenes →

Quick start

Linux · Python 3.12 · NVIDIA GPU with at least 28 GB free VRAM. First use downloads about 11 GB of base weights. Run with an HF login that has access to this repository:

python3.12 -m venv .venv
source .venv/bin/activate
pip install -U huggingface_hub
hf download IrohaStudio/Iroha-TTS --local-dir Iroha-TTS
cd Iroha-TTS
pip install -r requirements.txt

python infer.py \
  --text "やっと会えたね！ずっと待ってたんだよ。" \
  --emotion joy --intensity -0.015 --seed 7 \
  --ref my_voice.wav --ref-text "Exact transcript of the reference recording" \
  --out joy.wav

Replace my_voice.wav and the reference text with your recording and its accurate transcript. Output is 44.1 kHz WAV. Change --seed for another take; without a reference, the voice may also change with the seed.

For more details, ask your agent to read AGENT_GUIDE.md. It covers pinned downloads, Python / JSONL usage, parameter combinations, configuration and troubleshooting.

Parameters

Parameter	Default	Purpose
--text / --jsonl	Choose one	Single line / batch input
--emotion	none	joy, anger, sadness, surprise; neutral / none add no emotion tag
--intensity	Unset (α=0)	Four presets or an in-range float; shipped range [-0.06, 0.045]
--ref / --ref-text	None	Reference recording and accurate transcript, supplied together
--seed	0	Choose another integer to explore a different take
--guidance	Whisper 1.5, others 0	Vocal-event guidance; 0 disables it; separate from emotion intensity
--max-frames	400	Generation limit, about 18.6 seconds; no automatic sentence splitting
--no-argmax	Off	Also sample residual codebooks
--out / --out-dir	out.wav / outputs	Single output file / batch output directory
--model	Bundled model/	Model configuration and weights
--base	Cache/download	Optional local Fish S2-Pro base directory
--device	cuda	Inference device

Presets: weakest=-0.06, weak=-0.03, default=0, strong=0.045. Floats such as -0.015 are used directly, without snapping to a preset. Nonzero intensity needs an explicit choice of one of the four emotions and text without emotion or vocal-event tags. Use default intensity for transitions, laughter, sighs, crying and whispers.

Full parameter rules and configuration → AGENT_GUIDE.md

With emotion (joy, anger, sadness, surprise)

Model	Accuracy emo2vec large	Accuracy EmoRuler	Joy / Anger / Sadness / Surprise % EmoRuler
Fish S2 Pro (original) Starting point, no fine-tuning	17.9%	6.0%	6 / 0 / 9 / 9
Ours Fish S2 Pro + SFT + reinforcement learning	85.6%	93.3%	91 / 94 / 95 / 93
Fish Audio Drama 3 Latest open-source Fish model · 35B MoE (3B active)	32.1%	21.2%	32 / 14 / 21 / 18
IndexTTS-2.5 Open source · bilibili	40.2%	30.6%	65 / 18 / 19 / 20
Irodori-TTS-v4-Large · text description Open source · Aratako	30.8%	20.6%	34 / 23 / 8 / 17
Irodori-TTS-v4-Large · emoji Open source · Aratako	20.0%	9.8%	17 / 2 / 8 / 12
Chatterbox Open source · Resemble	Cannot specify emotion with tags

With / Without emotion

Model	With emotion			Without emotion
	Speaker similarity	Audio quality UTMOS	CER	Speaker similarity	Audio quality UTMOS	CER
Fish S2 Pro (original) Starting point, no fine-tuning	0.685	3.51	5.6%	0.684	3.57	6.2%
Ours Fish S2 Pro + SFT + reinforcement learning	0.636	3.55	5.8%	0.694	3.78	4.8%
Fish Audio Drama 3 Latest open-source Fish model · 35B MoE (3B active)	0.628	3.26	7.0%	0.653	3.43	9.3%
IndexTTS-2.5 Open source · bilibili	0.612	2.26	6.9%	0.714	2.91	6.0%
Irodori-TTS-v4-Large · text description Open source · Aratako	0.649	3.03	6.0%	0.662	3.09	5.7%
Irodori-TTS-v4-Large · emoji Open source · Aratako	0.651	3.03	6.5%	0.662	3.09	5.7%
Chatterbox Open source · Resemble	Cannot specify emotion with tags			0.675	3.34	6.5%

Subjective Listening Test MOS (Japanese)

Model	Naturalness ↑	Emotion match ↑	Speaker similarity ↑	Overall score ↑	Win share ↑
Ours	3.84	4.20	3.52	3.85	31.7%
Fish Audio Drama 3	3.97	3.64	3.45	3.69	23.2%
Irodori-TTS-v4-Large · text description	3.61	3.95	3.32	3.63	28.1%
Fish S2 Pro (original)	3.75	2.93	3.72	3.47	12.1%
IndexTTS-2.5	2.35	3.49	2.81	2.88	4.9%

All three dimensions are rated by human listeners on a five-point scale (MOS). The overall score is the unweighted mean of the three dimensions. Win share: the five models are compared for each listener–sentence pair, with ties for first place splitting the win equally.

Pairwise Win Rates

Pairwise win rates for Ours, based on the overall score, with ties counted as half a win:

Comparison	Ours win rate ↑
vs. Fish S2 Pro (original)	69.3%
vs. Fish Audio Drama 3	58.0%
vs. Irodori-TTS-v4-Large · text description	58.7%
vs. IndexTTS-2.5	82.7%

Binaural post-processing

spatial/ places finished speech around the listener's head for headphones, as in the whisper demo: circling, sweeping or held at one spot. It is a separate step, not a model feature, and needs no GPU.

pip install -r spatial/requirements.txt
python spatial/binaural.py --in a.wav b.wav --start -90 --turns 2 --near 0.8 --out orbit.wav

TODO

• Support more emotion and paralinguistic tags.
• Support emotion intensity control with all tags.
• Release Qwen3-TTS, Ming-3.0-tts, and OmniVoice models trained with a similar approach.
• Provide a WebUI for deploying different models and add support for sglang-omni.
• Release the GitHub project.

Contact

For technical questions, please contact ake.speechai@gmail.com.

License

This model is licensed under the Fish Audio Research License, Copyright © 39 AI, INC. All Rights Reserved (LICENSE.md). It is a Derivative Work of Fish Audio S2-Pro: research and non-commercial use is free; any commercial use requires a separate license from Fish Audio (fish.audio, business@fish.audio). The Fish S2-Pro weights are not included here and are downloaded from Fish Audio's repository under the same license. What was changed is listed in NOTICE. The inference code written for this release (emotts/, infer.py) is also available under the MIT License (LICENSE-MIT). spatial/binaural.py is under the MIT License, and the measured HRIRs in spatial/hrir/ are by TH Köln under CC BY-SA 3.0 (details).