IrodoriTTS-v2 Reference Audio

IrodoriTTS-v2 is a zero-shot voice cloning TTS, which is a fancy way of saying "it learns the voice from a short clip you give it." This node is how you give it that clip. It doesn't synthesize anything itself - it reads an audio file, packs it into a small config object, and hands that to the nodes that actually speak.

The output is ref_audio_config, and it plugs into exactly two places: the ref_audio_config input on an IrodoriTTS-v2 Sampler (single line of text) or the ref_audio_config_a / ref_audio_config_b inputs on IrodoriTTS-v2 Dialogue TTS (two-character conversations). In the dialogue case you run two of these nodes, one per speaker, so each character keeps their own voice.

The inputs are simple enough that you'll set two of them and forget the third:

• ref_audio - a dropdown of audio files (it also accepts video) in your ComfyUI input folder. That's the constraint: the clip has to live in ComfyUI/input/ for the dropdown to see it. Drag a wav or mp3 in there and refresh.
• normalize_ref_audio (default off) - when on, the node normalizes the clip to −16 dB and raises low-volume recordings so a quiet sample doesn't produce a quiet voice. For a whisper-quiet source clip, this is the difference between "usable" and "muffled."
• max_ref_seconds (default 30) - how much of the clip gets used. IrodoriTTS doesn't need your whole audiobook chapter as reference; 10–20 seconds of clean speech is the sweet spot, and capping it here keeps load fast and the voice coherent. A clip that's longer than this is truncated.

Mechanically it's refreshingly boring: the node hashes the file contents (so it can tell when you've swapped the audio), validates the file exists, and builds a small dict - path, normalize flag, and the second cap - that becomes ref_audio_config. It's the packaging step between "a file" and "an input the TTS understands."

One genuinely nice touch the README won't tell you: the pack ships a tiny web extension for this node, so the node itself draws an inline audio player in the UI. You can audition the reference clip right on the node before you spend a generation on a voice that's wrong. That's rare in TTS nodes and worth using.

Voice quality notes, from actually running these things: the reference clip matters more than any slider in this pack. A clean, single-speaker clip with no music and no background chatter beats a "better" clip that has the TV on in the room. If the output voice sounds like the room rather than the person, that's the reference, not the sampler.

If you're new to the pack, install it once and reuse it across the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/kantan-kanto/ComfyUI-DialogueTTS
pip install -r ComfyUI-DialogueTTS/requirements.txt

then restart ComfyUI (or search "ComfyUI-DialogueTTS" in ComfyUI Manager). The main checkpoint won't auto-download - see the Model Loader node's page for that dance, and remember the codec and tokenizer pull themselves on first load.

One honest caveat: IrodoriTTS-v2 is a Japanese model. It'll read English text, but it's at its best reading Japanese, and the reference clip should match the language you're generating. Don't feed it a Japanese reference and expect flawless English narration.