Hurston-1 开源 TTS 支持南方口音与窄带电话模式

Accented English text-to-speech built for the phone, steered by a plain-language voice description.

Hurston-1 introduces NB-TTS (Narrowband TTS): text-to-speech made for the phone line, where most customer service still happens, even when the other side is an AI agent. It learned in part from real phone calls between our contributors and AI voice agents, and its phone mode delivers what a caller actually hears: 8 kHz G.711 audio, ready for Twilio and other telephony systems. Full 24 kHz audio is still the default.

Hurston-1 is a fine-tune of Maya1 (Apache-2.0) built by Destined AI to speak English accents that most TTS systems handle poorly: Southern American, Nigerian, West African, East African, Southern African, alongside broader American English. It is the first open model in Destined AI's ClarityVoice family.

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

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
tts = Hurston('.', check=True)          
wav = tts("Well hey there, thanks for calling.", describe('Male', 'adult', 'Southern American accent'))

The South got something to say

In 1995 André 3000 stood on stage at the Source Awards, in front of a crowd that booed OutKast, and said "the South got something to say." We think that line still fits.

Hurston-1 was built in Atlanta by a small, independent team. Since 2021 we've been recording consented, paid contributors from the American South and from across Africa. These are voices that most AI systems still flatten into generic American English. Nobody handed us a funding round to do this, so we built it ourselves. These voices belong in AI, and now they have an open model.

We named it for Zora Neale Hurston. She grew up in Eatonville, Florida, trained as an anthropologist, and traveled the South recording how Black Southerners actually talked: their stories, songs and speech, at a time when few institutions thought those voices were worth keeping. Our work is a small echo of hers. The name is a tribute; it does not imply any endorsement by her estate.

Hear the South

A few lines the way people actually talk around Atlanta. All synthetic; every voice is generated from a description.

More samples for every accent are in demo/ (descriptions and text in demo/demos.csv).

Voice descriptions

The model was trained on descriptions of the form Gender, age, accent, style, pace:

Pace labels are absolute speaking rates learned from the natural voices in training data: roughly slow is under 2 words per second, measured 2 to 2.6, natural 2.6 to 3.4, fast over 3.4. The model follows them approximately, not exactly. It still highly matches the behavior of the underlying dataset.

Each call samples a new voice that fits the description. Note: Hurston-1 does not clone voices and was not trained to reproduce any real person.

Built for the phone (NB-TTS)

Most voice AI is demoed in studio quality and then deployed over a phone line. Hurston-1 is built for where it will actually be heard:

• Trained on real calls. About 10 hours of the training audio (7,600 turns from 250 contributors) is phone-quality speech from real calls with full-duplex AI voice agents, alongside studio-style reading and conversation.
• Phone mode. --phone (or phone=True) filters the output to the 300 to 3,400 Hz telephone band, resamples it to 8 kHz and encodes it as G.711 mu-law (--codec alaw for A-law), the formats phone networks, SIP and Twilio use. It is a channel effect applied after synthesis, so --check still judges the full-quality take.
• Ready to stream. ulaw_encode() returns raw mu-law bytes, the payload Twilio media streams expect.

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)   

What people build with it

Some places to start. At Destined AI we use accented TTS like this to build broad, high-quality test data, and that is still the use we would point to first.

• Test data for speech recognition. Generate the same sentences in nine accents and see where your ASR, captioning or transcription system breaks. It makes a quick accent-coverage check or a regression set before a release.
• Simulated callers for voice agents and IVR. Pair Hurston-1 with an LLM that writes caller scripts (customers, patients, applicants) and run hundreds of synthetic calls through your agent before real people do. Use phone mode so the agent hears what it will hear in production. Good for catching where an agent mishears accents, names or numbers.
• Training data augmentation. Add accented synthetic speech to ASR, keyword-spotting or wake-word training sets where those accents are thin. Mix it with real speech instead of training on synthetic audio alone.
• Voice prompts and announcements. Phone menus (phone mode output drops straight into an IVR), appointment reminders, store and transit announcements, app onboarding, in a voice that sounds like the community you serve.
• Education and training. Listening practice across world Englishes for learners, and accent-familiarity practice for customer service, healthcare and call-center staff.
• Accessibility and read-aloud. Read articles, forms or messages aloud in a voice closer to home.
• Games, animation and media drafts. Background characters, temp voice tracks and podcast or audiobook drafts, labeled as synthetic.
• Research. Accent-fairness audits of speech systems, listening and perception studies, and benchmarks that cover voices most datasets leave out.

Not a fit yet: one consistent brand voice across sessions (each call is a new voice; see Limitations), cloning a specific person, or anything safety-critical without a human listening first.

Evaluation

Two questions: does it sound like the accent, and does it say the words?

Accent strength vs base Maya1. Base Maya1 documents five accents (American, Indian, Middle Eastern, Asian American, British). For a fair comparison we prompted it in its own documented description style, naming each accent, and compared it with Hurston-1 on the same 10 sentences and 18 voices (180 clips per model). An accent classifier (African vs American English, trained on real speakers) scored each voice group:

The classifier cannot tell Southern American from general American English; that accent was checked by listening.

Intelligibility. Whisper large-v3-turbo word error rate (WER) against the input text, numbers normalized (lower is better).

Everyday sentences (the same 180 clips as the accent test above, natural pacing, single take):

Harder sentences (30 unseen sentences × 17 voices, 510 clips):

The harder set covers reading, health, customer service, names and numbers, expressive and long sentences. Base Maya1 is very clear because it mostly stays in American English; Hurston-1 keeps nearly the same clarity while actually doing the accent.

Names and numbers are the hardest category (WER 0.17 with --check). Part of that is the recognizer misspelling African names such as Adaeze Okafor rather than the model mispronouncing them.

Speaking rate. Median 2.84 words per second with natural pacing, against 3.76 for base Maya1. Hurston-1 speaks a little more deliberately; use fast pacing for quicker delivery.

Blind listening. Testers heard the same sentence from base Maya1 and Hurston-1 in random order, without knowing which was which, and chose the one that sounded more like a real person with the requested accent. Hurston-1 was chosen 78% of the time and base Maya1 6%; the rest were rated a tie or neither (54 comparisons). Base Maya1 was never chosen for an African or Southern American voice; its few wins were on plain American English.

Training data

LoRA (rank 64) fine-tune of Maya1, 2,000 steps, about 190 hours of speech:

Accent labels come from each contributor's self-reported background, checked against an audio accent classifier and human listening; speakers whose audio and labels disagreed were relabeled or removed. Pace labels come from each clip's measured speaking rate. All contributors were adults. The full Destined AI corpus (1,800+ hours, recorded since 2021) is not included in this release; contact us about licensing it.

Limitations

• No voice consistency across calls. Each call is a fresh voice that fits the description. Long text is split at sentence ends into chunks of about 40 words, so the voice can change between chunks. Consistent named voices are planned for a later release.
• Names, dates and amounts are the weakest content. Use --check and listen before you ship anything important.
• The generic African English accent label is weaker (WER 0.26); use a specific region instead.
• English only. No child voices.
• Like other autoregressive TTS models it occasionally produces a garbled or overlong take, or the voice shifts mid-take; --check catches most of these.
• Output is leveled to the same loudness for every voice, and pauses longer than half a second inside a take are shortened.
• Phone mode changes the channel, not the speaking style: the model does not yet choose call-style delivery (turn-taking, backchannels) on request. See What's next.

What's next

Hurston-1 is the first step. What we are building toward:

• Telephony-native TTS. A model that learns to talk like a person on a call, not just sound like the phone line: a call style distinct from reading style, with natural turn-taking, "mm-hm"s and call pacing, learned from our recordings of real people talking with full-duplex AI agents.
• Consistent voices. The same named voice across every call and every sentence, for agents and brands.

Follow Destined AI on Hugging Face for the next release.

Responsible use

Do not use Hurston-1 to impersonate real people, deceive listeners, or commit fraud. Tell listeners when speech is synthetic. Audio in demo/ is synthetic.

Credits and licenses

• Model weights: Apache-2.0, derived from Maya1 by Maya Research (Apache-2.0).
• Audio codec: SNAC 24 kHz by Hubert Siuzdak (MIT), downloaded at run time. Phone mode uses ITU-T G.711 (public standard).
• Optional checking: Whisper large-v3-turbo by OpenAI (MIT) and ECAPA speaker embeddings from SpeechBrain (Apache-2.0), downloaded at run time.
• Training data: LibriTTS-R (Koizumi et al., 2023; CC BY 4.0), Mozilla Common Voice (CC0), and Destined AI's own corpus.

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}
}

作者发帖说明（HN）：

Hey HN, I’m open sourcing my Hurston-1 narrowband TTS model (Apache-2.0). It is very good with American English and does especially well with Southern American and Regional African English accents. Destined AI is based in Atlanta, and I built this because I’m amazed by what AI can do, and I wanted to help introduce and welcome more people to feel comfortable using and interacting with AI.

Since 2021, I’ve been collecting fully consented voice data from over 20,000 people, 1M audio recordings, and over 1,800 hrs of audio captured to date. For this model, I used a small part of the data, 136 hrs, to finetune Maya1 and expand it’s capabilities.

Hurston-1 is a narrowband TTS model that focuses the output on telephone use because that is still where most of these voices will be heard. I’ve used it to support customers such as Amazon and Salesforce and generated over half-mill revenue with it. Now I’m open sourcing it to help unlock opportunities for others as well.

Limitations: this model doesn’t maintain the exact same voice across prompts and doesn’t support voice cloning. For many use cases, this is fine but maybe not so - if you’re creating podcasts.

This is the first release - I’m already working on the next version with improvements.

I would appreciate, if anyone is interested in this space, download the model from huggingface, check it out and let me know any feedback. Also, open to connecting with others in this space as well.