Irodori Audio for Anki

An Anki add-on that creates natural-sounding Japanese sentence audio with Irodori-TTS, running on your own computer. It picks a female or male voice and an emotion for every sentence, writes the audio to your SentenceAudio field, and never touches recorded audio.

Tested with the Kiku note type (fields Sentence and SentenceAudio), Anki 26.9.3, Windows 11 and an NVIDIA RTX 4080.

What it does

• Menu in the card browser: Cards > Generate Sentence Audio (Irodori).
• Six voices (3 female, 3 male) and seven emotions: neutral, happy, sad, angry, scared, whisper, joking.
• Voice and emotion are chosen by simple rules from the sentence (speech style such as 俺, ぼく, だぜ, わよ; keywords; punctuation). No AI model besides the TTS is needed. When unsure it uses neutral and a pseudo-random voice.
• Safe by default: notes with recorded audio are skipped, your Sentence field is never changed, old audio files are kept, and a whole run can be undone with Edit > Undo. Nothing changes if the server is down.
• Runs in the background with progress and a cancel button.

1. Run the Irodori server locally

You need an NVIDIA GPU with about 6 GB of free VRAM (CPU works but is slow), uv, git and ffmpeg (for mp3).

git clone https://github.com/Aratako/Irodori-TTS-Server.git
cd Irodori-TTS-Server
uv sync --extra cu128          # NVIDIA CUDA 12.8. Use --extra cpu for CPU only
cp .env.example .env

Edit .env:

IRODORI_HOST=127.0.0.1
IRODORI_HF_CHECKPOINT=Aratako/Irodori-TTS-v4-Small

Leave IRODORI_DEFAULT_NUM_STEPS unset. Keep the host on 127.0.0.1 so other computers cannot reach the server.

Copy the six .wav files from this repo's voices/ folder into Irodori-TTS-Server/voices/. Their names (female_calm, female_bright, female_cool, male_calm, male_young, male_deep) are the voice ids the add-on uses.

Start the server (the first request downloads the model, about a minute):

uv run --no-sync python -m irodori_openai_tts --host 127.0.0.1 --port 8088

Or use server/start-irodori.ps1 (Windows) or server/start-irodori.sh (Linux/macOS). Check it at http://127.0.0.1:8088/health.

Measured on an RTX 4080: about 0.9 s per sentence and 4-5 GB of VRAM.

2. Install the add-on

Install from AnkiWeb (code: 2078760911), download irodori_audio.ankiaddon from the releases page and open it with Anki, or copy addon/irodori_audio/ into your Anki addons21 folder. Restart Anki.

3. Use it

1. Start the server.
2. In the card browser, select a few notes and listen to the result first.
3. Cards > Generate Sentence Audio (Irodori). A dialog shows how many notes get audio and how many are skipped.
4. Notes with new audio get the tag irodori.

Settings are in Tools > Add-ons > Irodori Audio > Config: field names, audio format, overwrite behavior, voices and emotion captions.

Tips and known limits

• Pronunciation is not controllable. Unlike VOICEVOX there is no kana to correct. Rare names and unusual readings can be wrong. Listen first.
• Laughs and breaths are not guaranteed. The model sometimes adds a stray cry or breath after a sentence. The add-on detects clips that are far too long and regenerates them with another seed.
• Whisper is the least reliable emotion, especially for the male voices.
• Emoji in the text gave breathing artefacts in some tests, so emotions are set with a caption instead. The only exception is whisper, which also wraps the text in 🤫 and …. This only changes what is sent to the model, never your note.
• Reference voices should be short (about 3 s). A 4 s reference made one voice keep talking after the sentence ended.
• Do not use real people's voices (audiobook narrators, actors) as references. The model's terms forbid cloning a person without consent. The bundled voices were made with the model's own Voice Design (a text description, no reference audio) and do not imitate anyone.

Licenses and credits

• Add-on code: MIT.
• Irodori-TTS, Irodori-TTS-Server and the model Aratako/Irodori-TTS-v4-Small are by Aratako (MIT; see the model card for usage restrictions). Generated audio carries an inaudible watermark.