Hurston-1 is an Apache-2.0 fine-tune of Maya1 that speaks Southern American and African English accents from a plain-language voice description — the first model in Destined AI's ClarityVoice family.
2026-10-25 · by Moe Ameen
Destined AI has published Hurston-1 (`destined/hurston-1`) on Hugging Face, the first open model in what it calls its ClarityVoice family. Per the model card, Hurston-1 is an accented-English text-to-speech model "built for the phone, steered by a plain-language voice description." It is released under Apache-2.0, and the card describes it as a fine-tune of Maya1 (maya-research/maya1), trained as a LoRA on roughly 190 hours of speech. The weights are a 3B-parameter BF16 checkpoint; the repo is public but gated behind an access-conditions accept.
Rather than cloning a specific person, Hurston-1 generates a fresh voice on each call from a short text description — gender, age, accent, speaking style, and pace. The card lists Southern American English alongside a set of African English variants (Nigerian, including Yoruba- and Igbo-influenced, plus West African, East African, and Southern African), as well as General American. It is English-only and ships no child voices. The default output is 24 kHz audio via the SNAC codec, and a dedicated phone mode emits 8 kHz G.711 (the standard telephone band, roughly 300–3,400 Hz) for voice-agent and IVR use.
Destined AI reports its own evaluations on the card: in 54 blind listening comparisons, Hurston-1 was preferred 78% of the time over the base Maya1 (6%); word error rate on everyday sentences is about 0.03, and on harder sentences 0.145 for a single take, dropping to 0.062 with the optional `--check` re-synthesis pass (which uses Whisper and speaker embeddings to retry bad takes). The card is candid about limits: there is no voice consistency across separate calls yet, long text is split into roughly 40-word chunks between which the voice can shift, and names, dates, and amounts are the weakest content. No official release date is stated beyond the 2026 citation, and the demos and benchmarks are the developer's own — treat the figures as a snapshot of the model card rather than independent results.
Here is a concrete way to use Hurston-1 today without pretending a WAV file is a finished post. Draft the script inside [Kompozy](/) first — a Text Post, a Blog Article, or a short video script written to your [Persona Brief](/glossary/persona-brief) so the wording, banned phrases, and audience are already right — then run that script through Hurston-1 in a Southern American or African English voice and lay the audio over footage or a still to make a clip. Upload that clip to Kompozy's [Clipped Shorts](/glossary/clipped-short), which cuts it to vertical and burns in word-synced captions. That caption step is the one Hurston-1 can't do and the one that matters most: on muted autoplay, an authentic accent is invisible without on-screen text.
From there the same idea becomes a feed, not a single clip. Kompozy fans the topic into a [Carousel](/glossary/hyperframes), quote graphics, per-platform text, a blog, and a newsletter, all in the same Persona Brief voice, then [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email behind a per-post review. Hurston-1 gives a regional voice to the words; Kompozy writes the words and ships them everywhere.
Yes. Per the model card, Hurston-1 is released under Apache-2.0 with open weights on Hugging Face (destined/hurston-1), though the repo is gated — you log in and accept its conditions to download the 3B-parameter checkpoint. It is a LoRA fine-tune of the Apache-2.0 Maya1 base model.
The card lists Southern American English, General American, and a range of African English variants (Nigerian — including Yoruba- and Igbo-influenced — plus West African, East African, and Southern African). It is English-only with no child voices, and you steer each voice with a text description of gender, age, accent, style, and pace.
No. It generates a new voice on each call from your description rather than cloning a real speaker, and the card notes there is no voice consistency across separate calls yet — consistent named voices are listed as a future plan. Destined AI also asks that synthetic speech be disclosed and not used to impersonate people.