Hebrew Support

[saaruni1000]

Please add a Hebrew model so we can use Hebrew locally! Please comment if anything is needed for that such as a dataset, server to train a model on or anything else you might need!

[thewh1teagle]

Hi @saaruni1000 I'm working on Hebrew tts, do you have dataset for it?

[synesthesiam]

A dataset is definitely needed 🙂

[thewh1teagle]

@synesthesiam

https://huggingface.co/datasets/thewh1teagle/ILSpeech

This is dataset that I published recently The phonemes are manual tagged by linguistic And the audio recorded in studio Also all the phonemes exists in English except (inverted R, Glottal stop, and large X like in Arabic)

And of course the g2p is here: https://github.com/thewh1teagle/mishkal

[GeorgeS2019]

https://github.com/rhasspy/piper/issues/792

[synesthesiam]

I'm attempting to train 2 voices based on ILSpeech now 🙂

[GeorgeS2019]

@synesthesiam

We need many different Hebrew Voices.

Thanks for your attempts

https://github.com/rhasspy/piper/discussions/795#discussioncomment-13319144

https://huggingface.co/thewh1teagle/phonikud-tts-checkpoints

[synesthesiam]

Unfortunately, my training failed 🙁 The Hebrew phonemization in espeak-ng seems almost completely broken. For example, this seems to work fine:

espeak-ng -v he --ipa=3 -q "קוֹתֶבֶת הִיא תּוֹפָעָה טִבְעִית יָפָה"
kotˈevet ˈijʔ tofˈaoʔ tˈivijt jˈafaʔ

But changing the sentence results in espeak-ng thinking there are English words:

espeak-ng -v he --ipa=3 -q "קֶשֶׁת בַּשָּׁמַיִם הִיא תּוֹפָעָה טִבְעִית יפָה"
kˈof ˈe ʃ ˈe (‍e‍n)hˈiːbɹuːʃˈɪndɒt(‍h‍e) θˈav vˈet ˈa (‍e‍n)hˈiːbɹuːdˈaɡɛʃ(‍h‍e) ʃ ˈa (‍e‍n)hˈiːbɹuːdˈaɡɛʃ hˈiːbɹuːʃˈɪndɒt(‍h‍e) mˈem ˈa jˈod ˈi mˈem ˈijʔ tofˈaoʔ tˈivijt jfˈaʔ

Any ideas?

[GeorgeS2019]

Follow GitHub projects of thewh1teagle to understand how he replaces espeak

[synesthesiam]

I looked at phonikud-tts. It looks great, but seems to require a significant number of dependencies to run.

[GeorgeS2019]

Need to understand how special is the Hebrew language...it is not just another European language grammar but much more and much interesting. .speak to an orthodox Jews ..this person could explain better.

[GeorgeS2019]

@thewh1teagle

I have seen that you have more improvement.

Is your progress ready for release as the first of more official Hebrew voices for the piper?

[thewh1teagle]

I have seen that you have more improvement.

Is your progress ready for release as the first of more official Hebrew voices for the piper?

@GeorgeS2019

I'm gonna publish the project in the upcoming week and then we can add it

@synesthesiam

Piper doesn’t need to include Hebrew G2P. it can simply require an external G2P module when used with Hebrew. I believe it’s time for Piper to be less opinionated, especially considering how complex the G2P space is. Piper doesn’t have to be a single all in one executable for every language and setup.

While I understand the value of an all in one binary for use cases like Home Assistant, we can still support external G2Ps without sacrificing simplicity for those environments.

In fact, this change could significantly simplify Piper’s core. See: https://github.com/thewh1teagle/piper-onnx

[thewh1teagle]

@GeorgeS2019

I released the paper of phonikud, I also improved the TTS model, now we can release v1. See phonikud.github.io Note that the voice trained on non commercial data I'm not sure how it will work currently with piper

[thewh1teagle]

just released another enhanced Hebrew model, based on StyleTTS2

https://github.com/thewh1teagle/phonikud-StyleTTS2-dockerized

[thewh1teagle]

FYI the models in Hebrew are for non commercial / academic use (see the license)

[GeorgeS2019]

This issue is closed. https://github.com/rhasspy/piper/issues/792