CIRCLE WITH A DOT

@rommix0 I doubt so too, but it might be worth a try. My idea was based solely on the fact that Hebrew and Arabic have similar writing systems, in which vowels aren't written out in most texts. So synthesizers kind of have to guess at which vowel works where, and I figured BeST might just be taking the lazy approach and expecting you to write out the vowels.

I suspect something similar might be going on with Japanese, where Kanji are not being read out properly because they can actually have multiple correct pronunciations depending on context, where Hiragana and Katakana do spell out specific sounds so a synthesizer would have a much easier time trying to figure those out. The only thing that sustains that claim is that the word "konnichiwa" is read out properly and that one is usually spelled in Hiragana as far as I'm aware.

Of course, again, this is all speculation and I might be completely wrong, but I'd like to think my thoughts hold a little bit of water haha

@rommix0 This may be a ridiculous idea and I might be entirely misguided, but have you tried giving the Arabic one characters with vowel diacritics?
E.G. Without diacritics: اكتب جملة قصيرة
With diacritics: أُكْتُبْ جُمْلَةً قَصِيرَةً

@rommix0 This is... Truly something else. Never thought I'd hear this voice speaking so many languages. A few of them seem slightly broken, I'm pretty sure the Hebrew one is just reading the names of the letters out loud, and the Japanese one is doing... Something, not quite sure what, maybe to do with the 3 writing systems they have? I'm curious to play with these for sure.

In other Thursday morning news, VoiceOver has decided that it will sometimes read * at the beginning of a line as "star" and sometimes as "asterisk". This might have something to do with VoiceOver's language switching or it might have something to do with VoiceOver sending stuff to Eloquence weirdly. Thank god VoiceOver is so transparent with its internals and respectful to its users and will let me diagnose this without much hassle. Ahahahahahah funny joke. It's okay though, because this is one step from all of our synthesisers being replaced with AI voices whose text-to-phoneme rules are governed by the designs of divine providence alone.

@Onj I mostly agree with you on the principle of not dismissing LLM code outright, though I do think the analogy might be slightly misguided/mischosen. It's less like using a washing machine and more like using a hypothetical sort of clothes vending machine that puts together and sews your clothes on-demand. Your clothes are already made when you wash them, and presumably you read the washing instructions on the label and set your washing machine to the right settings. So yes, you're trusting it to follow those settings and to not mess up your expensive clothes, but you're not really having it create anything and the settings are quite limited-scope. I do think there is ethical use of AI, I try to make responsible use of it, as much as its background is very problematic and we ought to be conscious of that, so I definitely agree with you. But I also know that the ratio of shitty to decent AI-coded projects is much, much higher than the ratio of disastrous to successful washing machine cycles. Hey, how many tokens worth of water does a washing machine cycle use? Now that's a thought!