African American English Text to Speech: The Honest Answer in 2026

The short answer: no AAVE-trained voice exists here

EasyVoice does not ship a voice model trained on African American English (AAE), also called AAVE. am_michael and af_jessica read AAVE-grammar text accurately and default to a prosody closer to AAVE rhythms than brighter General American voices — but neither is an AAVE-trained model, and AAVE-familiar listeners will hear the difference.

What the nearest voices honestly do

am_michael and af_jessica are EasyVoice's closest fit for AAE content, and it's worth being precise about what that means. Both handle AAVE-grammar input — habitual "be," copula deletion, double negation — as written, without flagging it as an error or "correcting" it, because they read text as given rather than imposing a grammar model on top of it. Their default prosody sits closer to AAVE's rhythms than the brighter, faster General American voices like af_aoede or af_bella. That's genuinely useful for some jobs. What it is not: a model trained on AAVE speech itself. The specific phonology — vowel realizations, consonant cluster reduction, the actual sound of AAVE as spoken — isn't something either voice reproduces, because neither was built from AAVE training data. A listener who speaks or is closely familiar with AAVE will hear a General American voice reading AAVE-grammar text, not an AAVE speaker.

For character work, dialogue, or any performance where the audience needs to hear the accent as genuinely authentic — audiobook narration with AAVE-speaking characters, dramatized dialogue, culturally specific storytelling — a synthetic voice from us or anyone else isn't the right tool. Hire a human voice artist who speaks the variety natively. That's not a hedge; it's the honest answer for that category of work.

The real answer today: consent-first cloning

If the voice a project needs belongs to a real person who's willing to be cloned, EasyVoice's Pro voice cloning is the practical path. It works from a 15–60 second audio sample, requires a signed consent attestation before enrollment, and every generated output carries an AudioSeal watermark. Pro accounts can hold up to 3 custom cloned voices. The framing matters: this is consent-first cloning — clone a voice whose owner has actually agreed to it, never someone else's voice without permission, and never as a stand-in for approximating a whole community or variety rather than an individual who consented. See voice cloning for how enrollment works.

Why we haven't shipped a dedicated AAVE model

Building an AAVE-trained voice the right way means securing informed consent from AAVE-speaking voice donors, making sure those speakers are credited and fairly compensated, and doing the work carefully enough to avoid the cultural-appropriation risk a rushed, uncredited model would carry. That takes real time to do properly, and we'd rather ship it right than rush an inauthentic model to market to fill a keyword gap. It's on our longer-term roadmap; it isn't available today, and this page will change the day it is.

For the rest of the American English catalog — 28 voices across a range of registers — see American English voices.