Add Sephardic IPA schema

[charlesLoder]

It could be cool to have Sephardic but with IPA characters

[johnlockejrr]

Very cool but... Sephardic which? Spanish, Italian, Maghrebi etc. Sephardic reading tradition, like the Ashkenazic is not an unitar tradition.

[charlesLoder]

Great question! And the answer is, it depends.

For context, the overall idea is to produce something closer to a transcription (à la the Tiberian schema) so it can be fed through a Text2Speech tool like Google's or Amazon's.

The trick is, however, that I have to find a language model that will accept IPA characters. Amazon has a generic Arabic model which is probably close enough.

I jotted this down a little while ago, but has fallen to the way-side for the moment.

[johnlockejrr]

Did you find a model that accepts IPA? I think of training one for Tiberian.

[charlesLoder]

Most of the TTS models accept IPA characters via SSML. The one with the best results (imo) was Azure's AI Speech.

Using SSML I could give it this using Tiberian pronunciation

<speak version="1.0" xmlns="https://www.w3.org/2001/10/synthesis" xmlns:mstts="http://www.w3.org/2001/mstts" xml:lang="he-IL" style="documentary-narration">
    <voice name="he-IL-AvriNeural">
        <prosody rate="slow">
            <phoneme alphabet="ipa" ph="və.ha.ˈa.ʀɛts">
                וְהָאָ֗רֶץ
            </phoneme>
            <break time="250ms"/>
            <phoneme alphabet="ipa" ph="ha.jə.ˈtaː">
                הָיְתָ֥ה
            </phoneme>
            <break time="100ms"/>
            <phoneme alphabet="ipa" ph="ˈto.huː">
                תֹ֙הוּ֙
            </phoneme>
            <break time="100ms"/>
            <phoneme alphabet="ipa" ph="va.ˈvo.huː">
                וָבֹ֔הוּ
            </phoneme>
            <break time="250ms"/>
            <phoneme alphabet="ipa" ph="və.ˈxo.ʃɛx">
                וְחֹ֖שֶׁךְ
            </phoneme>
            <break time="250ms"/>
            <phoneme alphabet="ipa" ph="al.pə.ˈneɪ">
                עַל־פְּנֵ֣י
            </phoneme>
            <break time="100ms"/>
            <phoneme alphabet="ipa" ph="tə.ˈhoːm">
                תְהֹ֑ום
            </phoneme>
            <break time="250ms"/>
            <phoneme alphabet="ipa" ph="və.ˈʀuː.ax">
                וְר֣וּחַ
            </phoneme>
            <break time="100ms"/>
            <phoneme alphabet="ipa" ph="ɛ.lo.ˈhiːm">
                אֱלֹהִ֔ים
            </phoneme>
            <break time="250ms"/>
            <phoneme alphabet="ipa" ph="mə.ʀa.ˈxɛ.fɛt">
                מְרַחֶ֖פֶת
            </phoneme>
            <break time="250ms"/>
            <phoneme alphabet="ipa" ph="al.pə.ˈneɪ">
                עַל־פְּנֵ֥י
            </phoneme>
            <break time="100ms"/>
            <phoneme alphabet="ipa" ph="ha.ˈma.jiːm">
                הַמָּֽיִם׃
            </phoneme>
        </prosody>
    </voice>
</speak>

But the results were never quite good enough.

Khan's book and Kantor's site have audio, but I'm not sure if there is enough to train a model.

But don't let my ignorance stop you!

[charlesLoder]

Also, see this issue for some more #87