Some questions for improvement

[marwanlhabti5-coder]

Good project and amazing speaker similarity with decent pronouncing, but i have some questions as a native arabic speaker

In arabic there is something called "Diacritics" and it some informations in each letter and by it you decide how you will prononce the word, in this project its not working and when trying it, it produces nonsense words is there any way to support this or a way to train the model to support this kind of informations

And where did you get the dataset? If you can provide the dataset i can train the model or other model

And are you Arabic btw? I wonder bc i see you speak Chinese

[SWivid]

Hi @marwanlhabti5-coder , yes i speak Chinese lol

You could check all details in our paper: https://arxiv.org/abs/2601.13802
In short, we filter out clean samples from noisy ASR datasets (most of them are open-source)

For diacritics, cuz ASR datasets generally lack these marks, our model may not see enough cases during training, so not support well explicit control with diacritics (could check the vocab e.g. https://huggingface.co/SWivid/Habibi-TTS/blob/main/Unified/vocab.txt#L2546-L2703).
We are also surprised that TTS can work somehow without full diacritized text.
If you wish to finetune or even train from scratch, see #2 .

[marwanlhabti5-coder]

Wow thanks for the fast response

I'm struggling to find a proper dataset that contains arabic Diacritics, can you please provide me with the dataset you trained the model on (modern standard arabic) so i can diacritics it myself and some of you settings during training please

[marwanlhabti5-coder]

And @SWivid I'm curious why you are Chinese and producing models for Arabic tts? Lol

[SWivid]

can you please provide me with the dataset you trained the model on (modern standard arabic)

https://huggingface.co/datasets/pain/MASC for MSA as in Tab.1&2

[SWivid]

Just consider it as something I made for some of my Arab friends 😉

[marwanlhabti5-coder]

@SWivid Any chance you could fine-tune Habibi-TTS to support Arabic diacritics? 🙏

I have a humble request: would you consider officially fine-tuning the model to support Arabic diacritics (harakat) as an optional feature? This would make the quality like 50x better for classical Arabic, and educational use. I tried training it myself on my RTX 3060 12GB but failed miserably as a beginner 😅 kept running into memory issues and dataset errors

I actually have a diacritized dataset ready to share. Since you guys have the multi-GPU setup and deep knowledge of the codebase, you could knock this out way faster than us mortals. Even just sharing a training recipe or a pre-processed dataset would be super helpful

Totally understand if you're busy, but thought I'd ask. Keep up the great work ❤️

[SWivid]

@marwanlhabti5-coder how many hours diacritized data do you have.
We may put this in our todo list. Cuz we are gonna have a stronger base model, so we may test also Arabic performance and finetune on that one.

[marwanlhabti5-coder]

@SWivid the dataset i have is 3 hours and 40 minutes but I'm not sure if it open source or no

However I will add diacritics to this data (https://huggingface.co/datasets/pain/MASC) and upload it to huggingface and send you