Sounds like your project is for southern American English presumably including African-American Vernacular English (AAVE), and for regional accents in the African continent specifically for the English language.
Despite being focused on English, your project could be of interest for their funding. It's probably worth reaching out to let them know about your work.
Thanks for the tip, I will look into it. We wanted to broaden the model to include voices and accents from regions I don't typically hear or they might sound comical or like a caricature
We as an industry should not be working or associating with Bill Gates anymore or anyone else that was associated with Epstein and his cohorts. I just assume anyone still working with these people is complicit and needs to be thoroughly investigated.
I cannot speak to the other accents, but my experience has been that American south accents are really poorly represented in mainstream Speach-To-Text models. If your STT model is pointed at end users, this can cause quite a quality drift when looking at eval vs prod performance depending on who your user base is.
I am not sure what terms you have collected the speach data on, but I am sure there are lots of people working with voice models that would be very interested in this as an eval set.
thank you.. and agreed - that's why we decided to work on this because the tech is fascinating, but the representation is far behind. We have commercial rights for the data - that allowed us to create eval datasets but also build models.
AAVE was an awful term, but "Regional African English" and "Afro-English" is even worse if you're trying to talk about black Americans.
Are you talking about the English of Africans in different parts of Africa? If so, are you including the voices of white and Mediterranean Africans?
I'm confused. As a black American with roots (going back a couple centuries) in the Ark-La-Tex, I would think our dialects would be covered under "Southern American," which makes me suspect that you were referring to me with all of those references to another continent.
Good question, those are all different accents we are pointing out this model does well with, but it also does well with White American accents that are usually represented well, not only the Black voices (including White southern voices).
We are calling out Southern American, and the regional African English (e.g. a person with a regional Nigerian (igbo) accent speaking english, or someone with a regional Kenya (Swahili) accent that is still speaking english). Those are all different but all represented well in this model.
Wow this is really cool. AAVE and Southern lexes have been underserved linguistically for a long time.
That said, my hunch is your Africa speicifc datasets might lossy, as entire regions of Africa are very linguistically diverse and someone who speaks Bambara and French will affect a different English accent than someone who speaks Maninka and English.
Yes, exactly. This version focuses on giving the ability to consistently represent those voices well. The future versions will be even more fine-grained where you can scope it even more detailed if you wanted.
For example, right now if you say you want a Southern American accent, you might get a voice that is a Black person from Atlanta or Rural Georgia in future versions you will be able to specify Atlanta or rural, etc.. same with the regional African accents.
I agree, but I wasted like 5 minutes parsing the title.
(I imagine that it has the errors and bad pronunciation of a tipical Latam person, that is different for example from the errors and bad pronunciation of a tipical Spanish person)
But tangentially related, for non-English languages of the African continent, the Gates Foundation are pledging some grants to improve AI access outside of English-speaking regions: https://news.bpdata.com/article/gates-foundation-pledges-1-b...
Despite being focused on English, your project could be of interest for their funding. It's probably worth reaching out to let them know about your work.
I cannot speak to the other accents, but my experience has been that American south accents are really poorly represented in mainstream Speach-To-Text models. If your STT model is pointed at end users, this can cause quite a quality drift when looking at eval vs prod performance depending on who your user base is.
I am not sure what terms you have collected the speach data on, but I am sure there are lots of people working with voice models that would be very interested in this as an eval set.
Are you talking about the English of Africans in different parts of Africa? If so, are you including the voices of white and Mediterranean Africans?
I'm confused. As a black American with roots (going back a couple centuries) in the Ark-La-Tex, I would think our dialects would be covered under "Southern American," which makes me suspect that you were referring to me with all of those references to another continent.
We are calling out Southern American, and the regional African English (e.g. a person with a regional Nigerian (igbo) accent speaking english, or someone with a regional Kenya (Swahili) accent that is still speaking english). Those are all different but all represented well in this model.
Those were not a catch all for black americans.
That said, my hunch is your Africa speicifc datasets might lossy, as entire regions of Africa are very linguistically diverse and someone who speaks Bambara and French will affect a different English accent than someone who speaks Maninka and English.
For example, right now if you say you want a Southern American accent, you might get a voice that is a Black person from Atlanta or Rural Georgia in future versions you will be able to specify Atlanta or rural, etc.. same with the regional African accents.
(I imagine that it has the errors and bad pronunciation of a tipical Latam person, that is different for example from the errors and bad pronunciation of a tipical Spanish person)