Hurston-1 TTS: an Apache 2.0 voice model for Southern American and African-accented English, built for phone calls

Hurston-1 TTS is a new open-weight text-to-speech model for accented English, released on Hugging Face by Destined AI, a small team in Atlanta, and shared on Hacker News on October 25, 2026. It's a voice AI model you steer with a plain-language description of the voice you want, and its accent options include Southern American English and several African English accents. It also has a phone mode that outputs the 8 kHz audio that call systems use. Unlike some recent voice models, its weights are downloadable: Mistral's Voxtral 3, for example, is API-only, as our Voxtral 3 story explains.

This story is based on the Hurston-1 model card on Hugging Face, the card for the Maya1 model it was built from, Destined AI's website and the Show HN post, checked October 27, 2026. All results below come from the model card and were reported by Destined AI.

The short version

• What it is: a TTS model for accented English, steered by a written voice description and tuned for phone calls.
• Who built it: Destined AI, an Atlanta team that sells speech and other datasets. The name honors the writer Zora Neale Hurston; the card says it doesn't imply any endorsement by her estate.
• License and size: Apache 2.0, about 3 billion parameters in BF16. Downloading requires a Hugging Face login and accepting conditions.
• Accents: American, Southern American, Nigerian (general, Yoruba and Igbo), West African, East African (general and Swahili) and Southern African English. It doesn't list African American English.
• Limits: no consistent voice between requests, no voice cloning, and weaker handling of names, dates and amounts.

Does Hurston-1 speak African American English?

Not according to its model card. The Show HN post's title said the model speaks "Southern American and Afro-English," which is easy to read as African American English, though the post's own text describes "Regional African English" accents. The card's accent list covers Southern American English and accents from Nigeria and West, East and Southern Africa, and it never mentions African American English. The only reference to Black speech on the card is its description of Hurston's own work recording how Black Southerners talked. If you need African American English specifically, the card makes no claim that Hurston-1 covers it.

What data was Hurston-1 trained on?

Hurston-1 is a LoRA fine-tune of Maya1, a 3.3-billion-parameter open TTS model from Maya Research released under Apache 2.0 in late 2025. Destined AI trained it for 2,000 steps on about 190 hours of speech:

• About 136 hours from Destined AI's own corpus: scripted reading, conversation and about 10 hours of phone calls with AI voice agents. The card says contributors were consented, paid adults recorded since 2021, and that the phone portion is 7,600 turns from 250 contributors.
• About 46 hours from LibriTTS-R, a public audiobook speech dataset under CC BY 4.0.
• About 10.5 hours of African-accented English from Mozilla Common Voice, under CC0.

Accent labels came from speakers' own reports, checked with an audio classifier and by listening. The training data itself isn't released: the card says Destined AI's full corpus is more than 1,800 hours and invites buyers to "contact us about licensing it."

That makes Hurston-1 open-weight rather than open source in the strict sense the Open Source Initiative uses, which asks for training code and detailed data information, as our open-weight vs open-source explainer sets out.

How does it work, and what can it do?

You give Hurston-1 text and a voice description built from gender, age, accent, style and pace, such as an older Southern American woman speaking slowly in a clear conversational style. Each request creates a new voice from that description. The default output is 24 kHz audio; phone mode filters it to the 300 to 3,400 Hz telephone band at 8 kHz and encodes it as G.711 mu-law, for use with Twilio, SIP or interactive voice response systems. An optional --check step uses OpenAI's Whisper and a speaker-verification model to retry bad takes.

The card lists intended uses including test data for speech recognition, simulated callers for testing voice agents, phone prompts, education, accessibility and read-aloud, and research on accent fairness. It says Hurston-1 isn't a fit for a consistent brand voice, cloning a specific person, or safety-critical use without human review, and it asks users not to impersonate real people or deceive listeners and to disclose that speech is synthetic.

How well does it work?

All of these results were reported by Destined AI on the model card; it doesn't say who ran the listening test, and no independent evaluation has been published.

• Accent: on an accent classifier, Hurston-1 scored higher than base Maya1 on every African accent group, for example 0.95 against 0.60 for Southern African English. The card says the classifier can't tell Southern American from general American English, so that accent was checked by listening only.
• Intelligibility: word error rate, measured by transcribing the audio, was 0.03 on everyday sentences against 0.02 for Maya1. On a harder set of 510 clips it was 0.145 in a single take and 0.062 with --check.
• Preference: in 54 blind listening comparisons, listeners preferred Hurston-1 78% of the time and Maya1 6%.

The sample sizes are small, and the classifier and test sets aren't published as benchmarks, so treat these as the builder's own indicators.

What are Hurston-1's limitations?

The card lists them plainly. The voice isn't consistent across requests, and long text is split into chunks of about 40 words, so the voice can change partway through. Names, dates and amounts are the weakest content, which matters for phone systems that read out appointments or balances. Takes are occasionally garbled or run long. Phone mode changes the audio channel but not the speaking style, and call-style turn-taking isn't supported yet. It speaks English only and has no child voices. For a different approach to voice models, see our story on Block's speech-only Dessa model.

What's confirmed, reported and uncertain

• Confirmed (model cards): the builder, Apache 2.0 license, Maya1 base and LoRA setup, data sources and hours, accent list, phone mode and stated limitations.
• Reported by the builder: all accent, error-rate and listening results; consent and payment of contributors.
• Uncertain: when the model was first made public (the card gives no release date; Hugging Face's records show the repository was created on October 15, 2026 and last updated on October 17), and who Destined AI's customers are; neither the card nor the company's website names any.