Adding Hebrew Support to Rhasspy

[synesthesiam]

Hi Shay,

I run the Rhasspy project, a free, completely offline, open source voice assistant for as many languages as possible. A user on the Rhasspy forums recently asked about adding Hebrew, and I found this page. Would you be interested in helping?

I'd be happy to train a Kaldi nnet3 model for Hebrew, given enough speech data (at least a few dozen hours, 100's would be best). Unfortunately, there's no Hebrew on Common Voice or OpenSLR (except the yes/no dataset). I did find CoSIH, but haven't determined if it's viable yet.

If you could help crowdsource speech data with an open license (maybe using the SVLM corpus?), I will do the rest. If at least one Hebrew speaker has a good microphone and would be willing, I have a fork of MozillaTTS that I've been using to train text to speech voices.

[thewh1teagle]

@synesthesiam

Note that Hebrew is usually written without diacritics but they important for phonemes
And even the text is diacritized still there's no way to fully direct mapping it to phonemes and it will require more complex rules
I think both are solvable only with neural network or at least combination of it.

As for data in OpenSLR there's now pretty good data called SaSpeech