Show HN: Hurston-1 open source TTS that speak Southern American and Afro-English
huggingface.coHey HN, I’m open sourcing my Hurston-1 narrowband TTS model (Apache-2.0). It is very good with American English and does especially well with Southern American and Regional African English accents. Destined AI is based in Atlanta, and I built this because I’m amazed by what AI can do, and I wanted to help introduce and welcome more people to feel comfortable using and interacting with AI.
Since 2021, I’ve been collecting fully consented voice data from over 20,000 people, 1M audio recordings, and over 1,800 hrs of audio captured to date. For this model, I used a small part of the data, 136 hrs, to finetune Maya1 and expand it’s capabilities.
Hurston-1 is a narrowband TTS model that focuses the output on telephone use because that is still where most of these voices will be heard. I’ve used it to support customers such as Amazon and Salesforce and generated over half-mill revenue with it. Now I’m open sourcing it to help unlock opportunities for others as well.
Limitations: this model doesn’t maintain the exact same voice across prompts and doesn’t support voice cloning. For many use cases, this is fine but maybe not so - if you’re creating podcasts.
This is the first release - I’m already working on the next version with improvements.
I would appreciate, if anyone is interested in this space, download the model from huggingface, check it out and let me know any feedback. Also, open to connecting with others in this space as well.
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.
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.
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 generally feel the same, but on the flip side…making proper use of some of his money is probably one of his ex-wife’s favorite things in the world to do.
Food for thought.
Also grok (musk), and the strong allegations of SA against altmann (openai)
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
Do you have a way to demo the voice without downloading and running the model? The Huggingface playback doesn't work.
You can try this link to the huggingface spaces page https://huggingface.co/spaces/destined/hurston-1
I also have a demo phone line 678.940.8800 with a few of the voices
Trying to access the wave files directly reveals: Access to model destined/hurston-1 is restricted and you are not in the authorized list. Visit https://huggingface.co/destined/hurston-1 to ask for access.
Tangentially-related: When YouTube auto-dubs now, I noticed that Spanish language get dubbed in a Spanish-inflected accent, and French language similarly (I think).
Is it likely that they are using a model like this? Or are they actually trying to voice clone the original and then translate to English? I'm not even sure if the few Spanish-accented English auto-dubs I've heard are the same voice or not.
(Actually, that reminds me of growing up in Italy, where for a long time all English movies were shown dubbed, not subtitled, and they had the same lead male voice for all male leads, and the same lead female voice for all female leads, across every movie. I can still hear those voices in my head now thirty years later...)
Seeing Atlanna inspire an open weight model was not on my 2026 bingo card.
lol and there is more to come... we have more data, and can create even more fine grained controls for the model outputs
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.
Those were not a catch all for black americans.
congrats, this is a very exciting release.
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.
Regional African English accents, so Nigerian, Kenyan, South African?
Yes those are the ones for this model that performed consistently well.
does this mean they can make a roadman language translator now
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.
Southern American? It's Argentine or Chilean?
Given the context of "American English" and "Atlanta", I'm assuming they're referring to https://en.wikipedia.org/wiki/Southern_American_English
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)
You're thinking of "South America" (parse as "South America + n", or "from South America").
Yea I tried to squeeze info into the title - could def be better. But no this is referencing the southern states in America.