Numbers normalisation

[zidsi]

Checks

• This template is only for research question, not usage problems, feature requests or bug reports.
• I have thoroughly reviewed the project documentation and read the related paper(s).
• I have searched for existing issues, including closed ones, no similar questions.
• I am using English to submit this issue to facilitate community communication.

Question details

FYI - I've trained model with no number to string (words) normalisation text preprocessing for bravity. It picked pronaunciation well as well as context conditioned "word form" - which for same "number form" might differ in target language.

I wonder if anyone did some ablation study on this aspect? Which text preprocessing is a "must", which "nice to have" and which could be safely skipped?

Would mixed training samples (word and number representation for same speech target) confuse the model or help?

[SWivid]

Assuming that real-world data is infinite, I feel most natural not to have any text frontends, since any text frontends introduce priors.

In a nutshell, mixed training allows the model to make context-dependent decisions (as you mentioned), as well as support for specified readouts (using word instead of raw numbers, for example).

In this case, text preprocessing is often necessary when there is not enough data or when the data distribution is skewed.

p.s. There is also a special problem here, which is that it is difficult for the model to learn text regularization on its own in multilingual cases (e.g. F5-TTS base models tend to pronounce Arabic numerals in English, and it is reasonable to assume that the English corpus containing Arabic numerals dominates the pre-training data).

[Siddhesh-420]

Quick question i was searching for better tts for Android to replace the default Google tts
But found this and i liked the demo voices very much but unfortunately i don't have PC 🥲 so can u recommend something like this quality for Android 🙏🏻

[SWivid]

@Siddhesh-420 try previous lightweight NAR designs, like StyleTTS2 based models (e.g. Kokoro distills a lot of high-quality data. But the performance of different languages can vary considerably).

[Siddhesh-420]

I looked into kokoro but it doesn't state how to use it in android phones.

[Siddhesh-420]

@SWivid Thanks For Replying Tbh i am not that techy😅

I have been looking for a tts engine for Android Phone to replace default Google TTS to get more natural and human sounds. can u recommend the one easy to install like apk that works on the whole system level like Default Google TTS engine

[SWivid]

I have been looking for a tts engine for Android Phone to replace default Google TTS to get more natural and human sounds. can u recommend the one easy to install like apk that works on the whole system level like Default Google TTS engine

I dunno then

[Siddhesh-420]

Ohh
thanks for reply😄