Gender mismatch in voices (could be fixed with AI)

In languages that have grammatical gender, there’s an issue with male/female voices being selected at random regardless of the gender of the speaker based on the grammar of the sentence.

I’ve seen people discuss this in the past, but nowadays this problem could easily be fixed with very little manual labour by having AI go through every sentence and tag it as “said by a man” or “said by a woman”, so there’s really no reason I can see not to do this (even if it weren’t 100% accurate, it would still be way better than 50%, and the errors could then be fixed manually in little time).

3 Likes

It would be a great experiment if somebody proficient with AI could throw a bunch of sentences in a few languages at it, asking the AI to classify them into “spoken by a male”, “spoken by a female”, or “gender-neutral”.

1 Like

@clairchiara what languages + collections are you currently playing? To @morbrorper’s point, we can try experimenting a bit.

1 Like

@mike Currently having this problem with Hindi.

asking the AI to classify them into “spoken by a male”, “spoken by a female”, or “gender-neutral”.

For this, you’d need an AI that actually understands what male and female are. Good luck with that…

1 Like

All the major multilingual models are able to classify Hindi sentences by gender of speaker, see for example this test I ran with ChatGPT 4o with 10 random sentences from Clozemaster:

@morbrorper And with a slightly different prompt, you can get a less verbose, more organized answer (this last one made a partial mistake by classifying the first sentence as spoken by a female rather than gender-neutral, however it’s not a full mistake since gender-neutral sentences can be spoken by a female. The other 9 sentences were classified correctly):

1 Like

Hi, I’ve noticed this problem (man’s voice for a sentence that is grammatically a woman speaking, and vice-versa) in Polish as well.

A very simple (no AI needed) technique that would solve a very large majority of them is this:

Most of the texts involve a man or woman using a past tense first-person verb, e.g. “Czytałem gazetę” or “Czytałam gazetę” = “I was reading a newspaper”. If you could easily add a simple programmatic check check whether any word in the text ends with “łem” (masculine) or “łam” (feminine), then use a man or woman’s voice respectively.

(This simple check might sometimes cause a “false positive” (because it’s some other word that happens to end in “łem” or “łam” rather than the main verb of the speaker, e.g. a text with a quotation like “Mary said “I was reading a newspaper”” or whatever), and so in some minority of cases it would needlessly force a particular sex of speaker when it doesn’t really matter, but no harm done in that case, since it doesn’t matter whether a man or woman is speaking!)

If it is easy to add a programmatic check like this, then you could similarly check for the presence of a word ending in “łbym” or “łabym” in the text: these are the (less frequently used) masculine and feminine endings for first-person subjunctive, e.g. “Czytałbym gazetę” = “I would have been reading a newspaper”.

This simple trick would not catch all such cases, but I do think it would catch a large number of them.

A little more sophisticated analysis would be needed to catch another common class of sex-specific texts of the form “Jestem zdrowy” vs “Jestem zdrowa” = “I am healthy”, with adjective -y ending for a man and -a ending for a woman. :confused: A (relatively) simple non-AI regex trick that might work for such cases is to check whether “jestem” appears, and if so, check whether any word ends in “y” (then make a man’s voice" or any word ends in “a” (then make a woman’s voice) (and if there exists a y-word and an a-word, then this trick fails…)

2 Likes