📝 Phoneme Text Normalizer

TTS models mispronounce accented and non-English text all the time. Feed F5-TTS a Polish sentence full of ś, ż, ł or a German Grüße and it often stumbles on characters it wasn't really trained to read. This node sits before your TTS engine and cleans the text up first - either by simplifying tricky characters or by converting the whole thing to phonemes the model can pronounce reliably. It's text-in, text-out; the audio engine never sees the messy original.

Fair warning the author gives, and it's worth repeating: this is an experimental feature. It helps in a lot of cases, but "test with your target language" is genuinely the advice - results vary by language and engine.

How it works

You pick a method and it transforms the text accordingly. The interesting mode is IPA Phonemization, which uses espeak to rewrite your text as International Phonetic Alphabet symbols - an unambiguous "here's exactly how to say it" representation. The lighter modes just normalize awkward characters (splitting accents off letters, or mapping to plain ASCII) without going full phonetic.

The inputs and outputs that matter

• text (STRING) - the text to clean. Wire it in ahead of your TTS Text node, or type directly.
• method - the four modes, in rough order of aggressiveness:
  • Pass-through - no change (a bypass for A/B testing).
  • Unicode Decomposition - splits special characters into base letter + diacritic. Gentle, safe.
  • IPA Phonemization - full phonetic conversion via espeak. The strongest fix, and the one that needs espeak installed.
  • Character Mapping - ASCII fallback for maximum compatibility when a model really only wants plain letters.
• language (default Auto-detect, plus ~15 named languages) - leave it on Auto-detect and it guesses from the character patterns; set it explicitly if the guess is wrong for short or mixed text.
• show_debug (default on) - prints what it did, so you can see the before/after. Handy while dialing this in.

The single output is normalized_text (STRING) - connect it into the text input of your TTS engine.

How to install it

Comes with TTS Audio Suite. ComfyUI Manager: search TTS Audio Suite, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/diodiogod/TTS-Audio-Suite.git
cd TTS-Audio-Suite
python install.py

then restart. For the IPA Phonemization mode specifically you need espeak available: on Linux/Mac that's the phonemizer package plus system espeak, and on Windows the suite uses espeak-phonemizer-windows. The pack's installer handles the Python side; if IPA mode errors, a missing system espeak is the usual culprit.

Common issues & troubleshooting

IPA mode errors or does nothing. espeak isn't wired up. On Linux install system espeak (sudo apt-get install espeak-ng) so the phonemizer has a backend; the other three methods don't need it, so use Unicode Decomposition as a fallback.

Pronunciation got worse, not better. It happens - this is experimental, and phonemizing text an engine already handled fine can hurt. A/B it: run the same line through Pass-through vs your chosen method and keep whichever sounds right. Not every language/engine combo benefits.

Auto-detect picked the wrong language. On short snippets or mixed text the detector can miss. Set language explicitly instead of relying on Auto-detect.

It's compatible with which engines? Since it just outputs plain text, it works ahead of any TTS engine in the suite - F5-TTS, ChatterBox, and the rest. It's most useful for the phonetically-tricky European languages in F5-TTS.