Convenient way to specify phonemes

[balazsbme]

Hi,

I have used the project to create an audiobook in Hungarian. I would find it quite useful to be able to specify to the speech generation a list of phonemes, to correct if a foreign name would be pronounced differently than the output language.

For example, if a hungarian text contains the name, "Thomas", in the audio output I would prefer that it is pronounced with Hungarian phonetics (i.e. similar to the English pronounciation): "Tomasz".

As I understand this feature would expose the possibility for the use to input phonetic translations in some format. Similar to how the math symbols are replaced and read out:

ebook2audiobook/lib/lang.py

Line 115 in 1bb4418

language_math_phonemes = {

What do you think? Would this be a useful feature for others too? Or is it something that we could expect a user preprocesses in the text by replacing all occurances in the source ebook format?

Thanks,
Balazs

[DrewThomasson]

I think we already have this included in our lib/lang.py file at

abbreviations_mapping = {
    "eng": {
        "Mr.": "Mister",
        "Mrs.": "Mistress",
        "Dr.": "Doctor",
        "St.": "Saint",
        "Jr.": "Junior",
        "Sr.": "Senior",
        "Ave.": "Avenue",
        "Blvd.": "Boulevard",
        "etc.": "et cetera",
        "vs.": "versus",
        "e.g.": "for example",
        "i.e.": "that is",
        "et al.": "and others"
    },ect...

Which is open to any contributions from the community on additions to them for specific languages

[DrewThomasson]

Is that what you were referring to?

[DrewThomasson]

@ROBERT-MCDOWELL

What do you think? Or should we make a new thing in the lib/lang.py specifically for this use case.

As technically these are not abbreviations? 🤔

[ROBERT-MCDOWELL]

usually this is something that a model manages like xtts which is multiple languages.
for vits/fariseq we have to load the new language everytime a word should be in other language which could take ages and certainly fragment the RAM after load/unload many times the model. I'm afraid the best would be you find an hungarian phonetic that sounds closer to the other language pronunciation and avoid a very risky load/unload process which to my experience I won't code.

[DrewThomasson]

So I guess this is closed then?