How to handle custom pronunciations (brand names / acronyms) in realTimeTTS?

[sienaiwun]

Hi, thanks a lot for this awesome project!

I have a question about handling uncommon or custom words in realTimeTTS.
For example, suppose I want the model to pronounce a made-up brand name like "Xyloq".
By default, the phonemizer may guess something like ZAI-LOK, but I’d actually want it to be pronounced ZAI-LOH.

My questions are:

Is there a way to override the phonemizer and provide custom phonemes directly?

Does realTimeTTS support SSML tags or some kind of user lexicon?

If not, what’s the best practice for real-time scenarios? Should I rely on spelling variants (like writing “Xyloh”) to guide pronunciation?

Thanks in advance for any advice!

[fless-lab]

Hey @sienaiwun

You can do this with kokoroTTS

[sienaiwun]

Thanks a lot for the suggestion!
KokoroTTS does look great, but since it doesn’t support voice cloning (or finetuning), that’s the main reason I’m using realTimeTTS instead.

The challenge I’m running into is that with realTimeTTS I can’t seem to define custom pronunciations for certain words. Do you have any recommendations or workarounds for handling this?

[fless-lab]

Yeah got u but readtimeTTS is not a TTS model itself if i'm not wrong..It implements exiting models in order to offer something complete and mixed.. Including kokoro..

So when you use realtimeTTS, which model are u using behind ?

[sienaiwun]

I’m using the StyleTTS model with realTimeTTS.