Hurwitz-1 Open‑Source Accented TTS Brings Southern US and African English to Phone‑Line AI

Takeaway

Hurston‑1 is an open‑source narrow‑band text‑to‑speech model that can synthesize Southern American, Nigerian, West‑African, East‑African, Southern‑African and broader American English in 8 kHz G.711 phone‑line quality, making it ready for Twilio‑based voice agents and accent‑fairness testing.

What is Hurston‑1?

Hurston‑1 is a LoRA fine‑tune of the Maya1 TTS model (Apache‑2.0) released by Destined AI. It introduces NB‑TTS (Narrowband TTS) – a synthesis pipeline that outputs telephone‑band audio (8 kHz G.711 mu‑law) directly, while still supporting full‑band 24 kHz output. The model was trained on ~190 hours of speech, including ~10 hours of real phone‑call recordings between contributors and AI agents, to capture the acoustic characteristics of telephony.

Core Features

  • Accented English synthesis – Supports nine explicit accent descriptors: Southern American, Nigerian (general, Yoruba, Igbo), West African, East African (including Swahili), Southern African, and generic American English.
  • Voice description API – Voices are generated from a plain‑language description (Gender, age, accent, style, pace). Each request samples a new voice that matches the description; the model does not clone any real person.
  • Phone‑mode output – --phone (or phone=True) filters the waveform to the 300‑3,400 Hz telephone band, resamples to 8 kHz and encodes it as G.711 mu‑law (or A‑law). The resulting bytes can be streamed directly to Twilio, SIP or other telephony platforms via ulaw_encode().
  • Built‑in quality check – --check runs Whisper large‑v3‑turbo transcription and ECAPA speaker‑embedding verification; it retries synthesis when the recognizer cannot follow the audio or the voice changes mid‑take.
  • Open‑source licensing – Model weights are Apache‑2.0 (derived from Maya1). Training data combines Destined AI’s proprietary accented‑English corpus, LibriTTS‑R (CC BY‑4.0) and Mozilla Common Voice (CC0).

How to Use Hurston‑1

pip install torch transformers snac soundfile scipy speechbrain
python hurston.py --model . --check \
    --accent "Nigerian English accent" --gender Female \
    --text "Good afternoon. Please take one tablet twice a day after meals." \
    --out hello.wav

# Phone‑mode example (8 kHz G.711 mu‑law)
python hurston.py --model . --check --phone \
    --accent "Southern American accent" --gender Male \
    --text "Thanks for calling, how can I help you today?" \
    --out call.wav

from hurston import Hurston, describe, ulaw_encode
tts = Hurston('.', check=True)
call = tts(
    "Hi, I'm calling about my appointment on Tuesday.",
    describe('Female', 'older', 'Nigerian English accent'),
    phone=True,
)
payload = ulaw_encode(call)   # 8 kHz mu‑law bytes ready for Twilio media streams

Evaluation Results

Accent fidelity

A custom African‑vs‑American accent classifier was run on 180 clips (10 sentences × 18 voices) for each model. Hurston‑1 achieved substantially higher African‑accent probabilities than the base Maya1 model, e.g., Southern African 0.95 vs 0.60, Nigerian 0.94 vs 0.82, while keeping the generic American accent low (0.02).

Intelligibility

Test set Model Mean WER Clips >20 % WER
Everyday sentences (180 clips) Maya1 0.02 –
Everyday sentences (180 clips) Hurston‑1 0.03 –
Harder sentences (510 clips) Hurston‑1 (single‑take) 0.145 17.1 %
Harder sentences (510 clips) Hurston‑1 (--check) 0.062 8.4 %

When the --check flag is used, WER drops to 0.062, making the model reliable even on complex health‑care or numeric utterances.

Per‑accent WER (hard set, --check)

Accent WER
American 0.021
Southern American 0.070
Nigerian 0.070
Nigerian (Igbo) 0.016
Nigerian (Yoruba) 0.078
West African 0.059
East African 0.043
East African (Swahili) 0.019
Southern African 0.053

Names and numbers remain the most error‑prone (WER ≈ 0.17), largely due to recognizer misspellings rather than synthesis errors.

Listening test

In a blind comparison, participants chose the voice that sounded most like a real person with the requested accent 78 % of the time for Hurston‑1, versus only 6 % for the base Maya1 model.

Practical Use Cases

  • Accent‑coverage testing for ASR – Generate identical scripts in nine accents to expose transcription bias before a product release.
  • Synthetic callers for IVR/voice agents – Pair Hurston‑1 with an LLM that writes caller scripts, stream the 8 kHz mu‑law audio to Twilio, and evaluate how the agent handles diverse accents, names and numbers.
  • Data augmentation – Add accented synthetic speech to training sets for keyword‑spotting, wake‑word detection or low‑resource ASR models.
  • Localized voice prompts – Deploy phone‑menu prompts, appointment reminders or transit announcements in a voice that matches the community served.
  • Education & training – Provide learners with listening material in various world Englishes; train customer‑service staff to understand regional speech patterns.
  • Research – Conduct fairness audits of speech systems, perception studies, or benchmark creation for under‑represented English accents.

Limitations

  • No consistent voice identity – Each utterance samples a new voice; long texts are split into ~40‑word chunks, potentially changing voice mid‑sentence.
  • Higher error on names, dates, amounts – Use --check and manual listening for mission‑critical content.
  • Generic “African English accent” label is weak – Specific regional labels give better fidelity.
  • English‑only, adult voices – No child or non‑English language support.
  • Phone mode only filters audio, it does not alter speaking style – Turn‑taking cues such as “mm‑hm” are not automatically inserted.

Future Roadmap

  • Telephony‑native style TTS – A model that learns call‑specific prosody (back‑channels, pacing) rather than merely applying a band‑pass filter.
  • Consistent named voices – Ability to reuse the same synthetic voice across multiple calls and sentences, useful for brand‑voice applications.

Community Reaction (Hacker News)

  • A commenter noted that Southern US accents are “really poorly represented in mainstream Speech‑To‑Text models,” highlighting the potential for Hurston‑1 to improve evaluation datasets.
  • Another user pointed out the relevance of the project to funding opportunities such as the Gates Foundation’s grants for African‑language AI, suggesting outreach could broaden impact.
  • Several participants expressed enthusiasm for an open‑weight model that captures both Southern American and African English varieties, a combination rarely seen in 2026.

Responsible Use Guidelines

  • Do not use Hurston‑1 to impersonate real individuals or to deceive listeners.
  • Always disclose that audio is synthetic.
  • The demo audio in the repository is synthetic and should not be presented as real recordings.

Citation

@misc{hurston2026,
  title  = {Hurston-1: Accented English Narrowband Text-to-Speech (NB-TTS) for the Phone},
  author = {{Destined AI}},
  year   = {2026},
  url    = {https://destined.ai}
}

Sources