Hurston-1: First Open TTS for Southern American and Afro-English Accents, Built for the Phone Line
Destined AI, an Atlanta-based startup, has released Hurston-1 on Hugging Face under an Apache 2.0 license. It is the first open-weight text-to-speech model purpose-built for Southern American English and several Afro-English accent families — accents that existing TTS systems, both open and proprietary, routinely flatten into generic American English.
The model is a fine-tune of Maya1, itself Apache 2.0, trained on approximately 190 hours of speech assembled from three sources: Destined AI’s own consented corpus of accented English recordings, LibriTTS-R (CC BY 4.0), and CommonVoice. The consented corpus includes recordings from over 20,000 paid contributors and roughly 10 hours of real calls between contributors and AI voice agents — the actual content the model is designed to reproduce.
Accent Coverage
The model accepts plain-language voice descriptions in the form Gender, age, accent, style, pace:
| Accent | Sub-variants |
|---|---|
| American accent | — |
| Southern American accent | — |
| Nigerian English accent | Yoruba, Igbo |
| West African English accent | — |
| East African English accent | Swahili |
| Southern African English accent | — |
Style options include clear conversational, casual conversational, and clear audiobook narration. Pace is either natural pacing or deliberate pacing. The model does not clone voices — output is always generated from a description rather than a reference recording.
NB-TTS: Built for the Phone
The model’s distinguishing technical feature is what the team calls NB-TTS (Narrowband TTS). Standard TTS systems output wideband audio at 24 kHz or higher. Phone calls, including AI-to-human calls running through Twilio and similar platforms, operate at 8 kHz with G.711 mu-law or A-law encoding. The gap between wideband synthesis and telephony output is usually bridged by downsampling, which can introduce audible artifacts.
Hurston-1’s phone mode outputs directly at 8 kHz G.711, matching what the caller hears without an additional conversion step. The --phone flag activates this mode; ulaw_encode() returns raw bytes for direct integration with Twilio media streams. Wideband 24 kHz output remains the default for non-telephony applications.
Accuracy Check
The model uses Whisper for accuracy verification on generation. The check=True flag in the Python API retries takes that Whisper cannot follow or where the speaker voice changes mid-output — a practical guard against garbled synthesized speech that can occur with accent-steered fine-tunes.
from hurston import Hurston, describe
tts = Hurston('.', check=True)
wav = tts(
"Thanks for calling, how can I help you today?",
describe('Male', 'adult', 'Southern American accent')
)Context
Destined AI has been recording and compensating contributors from the American South and across Africa since 2021. The team describes the project as a deliberate response to a gap in AI voice systems: regional and diaspora English accents that get compressed into a generic output that does not match how those speakers actually talk.
The model is named for Zora Neale Hurston — the Eatonville, Florida-born anthropologist who recorded Black Southern speech, stories, and song in the 1920s and 1930s when few institutions considered it worth preserving. The name carries no implied endorsement by her estate.
Hurston-1 is the first open model in Destined AI’s ClarityVoice family. The roadmap in the model card references expansion of accent coverage and integration with hosted telephony infrastructure. The release appeared on Hacker News on October 26 with 34 points and 15 comments.