IrodoriTTS

Irodori-TTS 500M v2 is a Japanese-focused voice-cloning TTS model. It runs via a local FastAPI wrapper server in servers/tts/irodoritts/.

Why a script instead of a plain pip install

A direct pip install git+https://github.com/Aratako/Irodori-TTS fails for two reasons:

1. Packaging bugs in upstream pyproject.toml — the license field uses a bare string instead of a PEP 621 table, and the configs/ directory at the repo root causes setuptools auto-discovery to reject the flat layout. The install script patches both before building.
2. dacvae is not on PyPI — the upstream repo declares it via [tool.uv.sources], which is a uv-only extension that pip ignores. The script pre-installs dacvae directly from GitHub before installing irodori-tts.

Setup (Windows PowerShell)

# 1. Create and activate a venv inside the engine folder
python -m venv scripts\tts\irodoritts\.venv
scripts\tts\irodoritts\.venv\Scripts\Activate.ps1

# 2. Upgrade pip
python -m pip install -U pip

# 3. Install server runtime deps (FastAPI, uvicorn, PyTorch)
pip install -r scripts\tts\irodoritts\requirements.txt

# 4. (GPU only) Reinstall PyTorch with CUDA support — skip for CPU-only installs
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu124

# 5. Install irodori-tts from source via the patch script
.\scripts\tts\irodoritts\install-irodori.ps1

# 6. Start the server
python scripts\tts\irodoritts\server.py

CUDA version: replace cu124 with cu118 or cu121 if your driver targets an older toolkit.

Security note

Both Irodori-TTS and dacvae are installed from GitHub. The install script pins both to specific commit SHAs (defined at the top of install-irodori.ps1) to prevent silent upstream changes from affecting installs. When updating, replace the SHA constants with the new HEAD commits and verify the diff before deploying.

Registering in TomoriBot

After the server is running, register it with /provider custom-endpoint add:

• capability: speech
• api_style: tts-clone
• endpoint_url: http://127.0.0.1:8013
• script_markup: emoji
• supports_instruct: false

Then select it with /model speech, upload a reference sample with /speech voice-add, and assign it with /speech voice-assign.

TomoriBot strips Discord custom emoji syntax before sending text to TTS. With script_markup: emoji, Unicode emojis are preserved for IrodoriTTS emotion control; other speech modes remove Unicode emojis too so they are not spoken literally.

Environment variables

Variable	Default	Purpose
IRODORI_TTS_MODEL_ID	Aratako/Irodori-TTS-500M-v2	HuggingFace model repo
TOMORI_TTS_HOST	127.0.0.1	Server bind address
TOMORI_TTS_PORT	8013	Server port
IRODORI_MODEL_DEVICE	cuda / cpu	Inference device
IRODORI_CODEC_DEVICE	same as model device	Codec device
IRODORI_MODEL_PRECISION	bf16 (GPU) / fp32 (CPU)	Model precision
IRODORI_CODEC_PRECISION	fp32	Codec precision
TOMORI_TTS_MAX_TEXT_CHARS	1000	Per-request text length cap