Destined AI released Hurston-1, an open TTS model for African English accents: Reveneau

Why it mattersVoice agents often cannot speak the accent of the customer on the line. A permissive open-weight model trained on consented voices from the American South and Africa closes part of that gap.

A voice agent that answers a Lagos customer in a Lagos accent is still rare in a production stack. Destined AI, a small team in Atlanta, has posted Hurston-1 on Hugging Face. The model is a 3-billion-parameter open text-to-speech model under the Apache-2.0 licence. The Show HN launch for it reached 45 points on Hacker News.

What the model covers

Destined AI says Hurston-1 is a LoRA fine-tune of the Maya1 base model, at rank 64. The accents on the model card include American and Southern American. The African list adds Nigerian, Nigerian Yoruba and Nigerian Igbo English. It also adds West African, East African, East African Swahili and Southern African English. The company says each voice is picked from a short written description. A user writes a gender, age, accent, style and pace, and Hurston-1 produces a fresh voice from that line.

What Destined AI reports for quality

Destined AI reports a median word error rate of 0.03 on everyday sentences. The company says harder sentences score 0.145 on a single take. The same sentences score 0.062 with a Whisper-based retry flag turned on. The company says it ran a blind listening test of 54 side-by-side pairs. Listeners picked Hurston-1 in 78 percent of pairs and the Maya1 base model in 6 percent. Destined AI's own accent classifier scores the African voices between 0.84 and 0.98 African probability. The American voices score under 0.02 on that same classifier. Every one of these numbers is the company's own, and no outside party has repeated them yet.

The training set, in Destined AI's words

The company says it has recorded voice contributors in the United States and Africa since 2021. The pool holds 1 million audio clips and over 1,800 hours of audio. For this release, the company says it used 136 hours from that pool. It also added 46 hours from LibriTTS-R and 10.5 hours from Mozilla Common Voice. About 10 of those 136 hours are real phone calls between live callers and a voice agent. The company says every contributor consented and is credited in the NOTICE file.

The phone mode and the limits

Hurston-1 includes a narrowband mode built for a phone system. Destined AI says the output is 8 kilohertz G.711 mu-law audio, the format a Twilio phone bot expects. The model card also lists honest limits. The voice is not consistent across calls. A brand voice that returning customers recognise is not what Hurston-1 does yet. Names, dates and amounts have higher error rates than ordinary speech. The model is English only, and there are no child voices.

Many voice-agent teams serve customers in an accent a big hosted API does not cover. Those teams have had two bad options. The team could record its own voice set. That is slow and costly. The team could pick a generic American voice. Many callers do not recognise that voice as theirs. Hurston-1 is a third option under a permissive licence. The cost is that only Destined AI has measured it so far.

Source

Destined AI: destined/hurston-1 on Hugging Face
Launch discussion: Hacker News thread

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.