Hurston-1 open source TTS that speak Southern American and Afro-English

I founded Destined AI, a small Atlanta company that has collected over 1,800 hours of consented, paid speech recordings from Southern American and African English speakers since 2021.

Most TTS models handle accents through the prompt. Ask for a Southern or Nigerian accent and you get something close to American English - which is good for may use cases. However, we fine-tuned Maya1 (3B, Apache-2.0) with LoRA on about 190 hours of speech, 136 of them from our own corpus, and we are releasing the weights. Sometimes we want voices that sound more regional and representitive.

We built it for phone calls, because that is where most voice agents and IVRs are actually heard. There is a phone mode that outputs 8 kHz G.711, the format Twilio and SIP carry, and a good portion of the training audio is real calls between our contributors and AI voice agents. We are calling the category NB-TTS, narrowband TTS.

This is the first release - more to come.