Gerald Carter (@LGcarter) open-sourced Hurston-1, a TTS model for accented English built with phone-based voice agents in mind. He founded Destined AI after seeing AI struggle with Southern US accents. Since 2021, his Atlanta team has recorded paid, consenting speakers from the American South and across Africa. "Nobody handed us a funding round to do this, so we built it ourselves." The model is named after Zora Neale Hurston, the writer and anthropologist who traveled the American South recording Black communities' stories, songs, and speech. Hurston-1 is a 3B-parameter LoRA fine-tune of @mayaresearch_ai's Maya1 on ~190 hours of speech, including ~10 hours of real voice-agent calls. - Voice by description — gender, age, 9 accents, style, and pace. - Phone mode — post-synthesis 8 kHz G.711 audio for Twilio. - Voice-agent testing — generate accented callers to find where speech recognition and agent workflows struggle. In 54 blind comparisons, listeners preferred Hurston-1 over base Maya1 for sounding like a real person with the requested accent, 78% to 6% (team-reported). Voices can vary between generations, so it's better suited to synthetic caller testing than maintaining one consistent voice. Open weights under Apache-2.0. Training corpus not released. huggingface.co/destined/hurs… huggingface.co/spaces/destin… #BackchannelSignals