Destined AI open-sources Hurston-1, a TTS model for Southern American and African English accents

Destined AI, a small independent team based in Atlanta, has released Hurston-1, an open text-to-speech model built for English accents that mainstream systems tend to smooth into a default American voice: Southern American English alongside Nigerian, West African, East African and Southern African varieties. According to the model card published on Hugging Face, which reached Hacker News's front page on 25 October 2026, the release also introduces what the team calls NB-TTS: synthesis designed for the phone network rather than the studio.

How voices are specified

Hurston-1 is a fine-tune of Maya1, an Apache-2.0 speech model, and is the first open release in Destined AI's ClarityVoice family. Rather than cloning a reference speaker, users describe a voice in plain language: gender, age, accent, style and pace. Nine accents are supported — American, Southern American, Nigerian English, Nigerian (Yoruba), Nigerian (Igbo), West African, East African, East African (Swahili) and Southern African — with styles covering clear and casual conversation plus audiobook narration. Pace labels correspond to absolute speaking rates measured from the training voices, from under two words per second up to more than 3.4. Each call samples a new voice matching the description; the card states plainly that the model does not clone voices and was not trained to reproduce any real person. An optional check mode re-records takes that a Whisper-based transcriber cannot follow, or where the voice drifts mid-clip.

Narrowband synthesis for the phone line

The central pitch is that voice AI is usually demonstrated in studio quality but deployed over a telephone channel. Hurston-1's default output is 24 kHz audio, but its phone mode applies the 300–3,400 Hz telephone band, resamples to 8 kHz and encodes as G.711 mu-law (A-law is also supported), the formats used by phone networks, SIP trunks and platforms such as Twilio. Because the effect is applied after synthesis, the quality check still judges the full-band take, and a helper function returns raw mu-law bytes ready for Twilio media streams. Training reflects the channel as well: roughly 10 hours of phone-quality speech, about 7,600 turns from 250 contributors, came from real calls between the project's contributors and full-duplex AI voice agents, mixed with studio-style reading and conversation.

The team and the name

Destined AI says it has recorded consented, paid contributors from the American South and across Africa since 2021, and built the model without outside funding. The name honors Zora Neale Hurston, the writer and anthropologist who traveled the early twentieth-century South recording Black Southern speech; the card is explicit that the tribute implies no endorsement by her estate. The team frames the work with a nod to André 3000's 1995 Source Awards remark that "the South got something to say."

Evaluation and limitations

For evaluation, the team compared Hurston-1 against base Maya1 — which documents five accents of its own (American, Indian, Middle Eastern, Asian American and British) — on identical text and voices, 180 clips per model, scored by an accent classifier trained on real speakers. The results table is cut off in the publicly available card text, so the specific scores are not reproduced here. The card is candid about fit: each call produces a new voice, so a consistent brand persona across sessions is out of scope, as is cloning a specific person or any safety-critical use without a human listening first. Suggested applications center on testing: generating the same sentences in nine accents to expose weak spots in speech recognition, running simulated accented callers through voice agents in phone mode before real people call, and augmenting ASR or wake-word training data. IVR prompts and announcements, listening practice across world Englishes, read-aloud, media drafts labeled as synthetic and accent-fairness research round out the list.

Why it matters

Accent coverage is a persistent weak spot in speech technology, and open tooling to probe it is scarce. The model card argues that these speakers are usually flattened into generic American English, which leaves voice agents and recognition systems untested against a large share of real callers. An openly available generator for nine accents gives teams a practical way to build accent-coverage regression sets and simulated call traffic before deployment. The phone-first design matters just as much: an AI voice agent ultimately hears an 8 kHz G.711 channel, not studio audio, and Hurston-1 lets developers test against exactly that. Independent evaluation will determine whether the quality holds up, but the gap this project targets — accented English over phone lines — is real and largely unaddressed by open models.