All posts
Text to Speech

Kokoro 82M Voices: All 11 Graded, and Which to Pick

Eleven voice IDs, zero descriptions. We put Kokoro's official grades next to every name, ran one line through all 11 on each::labs, and matched each voice to a job.

By Eftal Yurtseven · Nov 17, 2026 · 9 min read

Kokoro 82M Voices: All 11 Graded, and Which to Pick

Eleven names in a dropdown, and not one of them tells you what you're getting. af, af_bella, bm_lewis. You pick one, render a paragraph, shrug, pick another, and half an hour later your product video still has no voice. That's the real friction with Kokoro 82M voices: the model is quick, the choice isn't.

This guide fixes the choice. Kokoro 82M on each::labs exposes exactly 11 voices through its voice field, and Kokoro's own voice card on Hugging Face grades the training data behind each of them. We put the grades next to the names, ran the same line through all 11 voices on the each::labs API, and matched each voice to a job. If you want the model overview first (what Kokoro is, why a small model wins often), read our Kokoro 82M spotlight. This post is only about the voices.

1. Read the Name Before You Press Play

Every Kokoro voice ID is a tiny spec sheet. The first letter is the accent, the second is the gender, and the rest is the speaker's name. Kokoro's original model card spells out the first part: language is determined by the first letter of the voice name, where a means American English (en-us) and b means British English (en-gb). The voice card marks the second letter's meaning with its female and male icons, so f is a female voice and m a male one.

So bm_george reads as British, male, George. af_sky is American, female, Sky. Once you see it, the dropdown sorts itself into four groups: seven American voices (the default blend, four named female voices and two male) and four British voices (two female, two male).

The 11 IDs on each::labs are the exact lineup of Kokoro's first public release, v0.19, which hexgrad published between December 25 and 31, 2024: Bella and Sarah first, then Adam and Michael, then the four British voices, then Nicole and finally Sky. Hexgrad's current release lists 54 voices across 8 languages, and its VOICES.md card is where the grades below come from.

A name is a spec sheet if you know how to read it.
A name is a spec sheet if you know how to read it.

2. Know What the Grades Actually Measure

Don't read the card's letter grades as a beauty contest. Hexgrad says plainly that they estimate the quality and quantity of each voice's training data, and that voices will sound better or worse to different people. Two inputs feed the overall grade. Target quality asks how clean the reference audio was and how well the text labels matched it. Training duration asks how much of that voice the model heard, in tiers from a few minutes to tens of hours.

That second number matters more than people expect. When you're deciding which voice to trust with a 2,000-word script, the duration tier is the closest thing to a reliability score you'll get.

Grades measure the data, not the charm.
Grades measure the data, not the charm.

3. The Kokoro 82M Voices at a Glance

Here's the whole set in one place, with the card's grades and what we measured. Our test line was the same 111 characters for every voice: "Your order has shipped. It should arrive on Thursday, and you can track every step from the link in this email." Speed was left at the default of 1.

voice        accent    gender   card grade   training data   our clip
af           American  female   not graded   Bella + Sarah   7.17 s
af_bella     American  female   A-           10 to 100 h     7.17 s
af_nicole    American  female   B-           10 to 100 h     10.35 s
af_sarah     American  female   C+           1 to 10 h       7.20 s
af_sky       American  female   C-           1 to 10 min     6.58 s
am_michael   American  male     C+           1 to 10 h       7.47 s
am_adam      American  male     F+           1 to 10 h       6.47 s
bf_emma      British   female   B-           10 to 100 h     6.47 s
bf_isabella  British   female   C            10 to 100 min   6.47 s
bm_george    British   male     C            10 to 100 min   7.20 s
bm_lewis     British   male     D+           1 to 10 h       7.10 s

Every clip came back as a mono WAV at 24 kHz (32-bit float), and each prediction finished in roughly 2 to 7 seconds of processing time. Most voices landed between 6.5 and 7.5 seconds for the line. The outlier is af_nicole at 10.35 seconds, which tells you something useful before you've heard a note: Nicole takes noticeably longer over the same words, so budget for that in timed slots.

4. Start With af or af_bella for American Narration

The default voice, plain af, isn't a person. Kokoro's v0.19 card describes it as a 50-50 mix of Bella and Sarah, and notes that this blended voice is the one that was ranked in the TTS Spaces Arena when Kokoro launched. It's what you get if you send no voice at all. That makes it a reasonable baseline, and a reason to test the two voices it's made from.

af_bella is the strongest entry on the card for this set: an A target quality, the tens-of-hours training tier and an overall A-. No other voice on the each::labs endpoint grades higher. If you only have time to test one voice for long American English narration (audiobook chapters, article read-alouds, course modules), make it Bella. Then run the same paragraph through af and keep whichever your listeners prefer.

Pick the accent first. The speaker follows.
Pick the accent first. The speaker follows.

5. Give af_nicole, af_sarah and af_sky Their Own Jobs

af_nicole is the other American female voice with tens of hours behind it, graded B-. Her slower read of our test line makes her a sensible pick for content where unhurried delivery helps, such as onboarding walkthroughs, meditation-style app copy or explainers people play while they work. If she runs long for a slot, nudge the speed up rather than cutting words.

af_sarah sits at C+ with 1 to 10 hours of data, and she's the other half of the default blend. Treat her as your alternate American voice: useful when one product needs two distinct female speakers, say a host and a support agent in the same demo.

af_sky is the thinnest voice in the set. The card puts her training data in the 1 to 10 minute tier and grades her C-. Use her for short, low-stakes lines, prototypes and placeholder audio, and do a full listen before she carries anything long.

6. Pick am_michael Over am_adam for American Male

There are only two American male voices here, and the card doesn't treat them equally. am_michael has a B for target quality, 1 to 10 hours of data and an overall C+. am_adam has the same duration tier but a D for target quality, which drags him to F+, the lowest grade in the set. Hexgrad's note on target quality says misaligned labels or weak reference audio pull that score down, so this is about the data he was trained on, not a matter of taste.

Our call: am_michael is the default male voice for American explainers, product tours and narration. Keep am_adam for when you need a second male character, and listen to his output closely before shipping.

Pace is a dial. Length is the other one.
Pace is a dial. Length is the other one.

7. Choose bf_emma First for British English

British English is where Kokoro 82M voices earn their place in a lot of pipelines, because it's a real accent switch on the same request, not a prompt trick. bf_emma leads the British side: B target quality, the tens-of-hours tier and an overall B-, the only British voice in this set with that much training audio. For UK-facing narration, e-learning and documentary-style reads, she's the first test.

bf_isabella grades C with 10 to 100 minutes of data, a solid second British female voice for dialogue or a contrasting host. On the male side, bm_george also grades C with the same minutes tier, and bm_lewis has more hours but a C target quality, landing at D+. Start with George for British male narration and keep Lewis as the alternate.

8. Which Kokoro 82M Voice for Which Job

Put the grades and the accents together and the choices get simple. For long-form American narration, test af_bella and af. For long-form British narration, bf_emma. For an American male explainer or product tour, am_michael. For a British male voice, bm_george.

Ads are different. They're short, and Kokoro's voice card says most voices perform best on 100 to 200 tokens (out of roughly 500 possible) and can sound weaker on very short utterances, especially under 10 to 20 tokens. A four-word tagline on its own is the worst case. Write the tagline into a full sentence or two of the spot, render it together, and trim in the edit. For the voice itself, pick from the higher-graded group (Bella, Nicole, Emma) so the short read starts from the strongest data.

For dialogue, pair across groups so the speakers never blur: af_bella with am_michael for an American scene, bf_emma with bm_george for a British one, or mix accents when the characters come from different places. For prototypes and placeholders, anything goes, which is where af_sky and am_adam fit best.

9. Set Speed and Chunk Size Before You Blame the Voice

The speed field on each::labs is a multiplier: 1 is the default, 0.5 is half speed and 2.0 is double. We re-ran the test line at 1.25 for two voices. af_bella went from 7.17 to 6.00 seconds and bm_george from 7.20 to 6.17. That's about 14 to 16 percent shorter, not the full 20 percent you might calculate from the multiplier, so measure the real file before locking a timed edit.

Length is the other dial. The text field accepts long input and splits it into smaller chunks for you, but the voice card warns that voices can rush on long utterances, especially over about 400 tokens, and suggests chunking or adjusting speed. In practice, write in paragraphs that sit in that 100 to 200 token range. As a rough guide, Kokoro's card says 1,000 characters of input is about one minute of audio on average, and our 111-character line came out at about 6.5 to 7.5 seconds for most voices, right in line with that.

10. Send One Request, Swap the Voice

Kokoro 82M runs through the same prediction API as every other model on each::labs. Create a prediction with the model slug and your three inputs:

curl -X POST https://api.eachlabs.ai/v1/prediction/ \
  -H "X-API-Key: $EACHLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kokoro-82m",
    "version": "0.0.1",
    "input": {
      "text": "Your order has shipped. It should arrive on Thursday, and you can track every step from the link in this email.",
      "voice": "bf_emma",
      "speed": 1
    }
  }'

The response carries a predictionID. Poll GET https://api.eachlabs.ai/v1/prediction/<id> until the status is success, and the WAV URL is in output. If you'd rather not poll, add a webhook_url to the create request. The fastest way to choose: loop over all 11 IDs with your real script line and listen side by side. To chain the audio into a video or lip-sync step, build it as a flow in each::labs flows.

11. When 11 Voices Aren't Enough

Kokoro on this endpoint is English only, preset voices only, and no cloning. When a job needs more, stay on the same API and change the model. Inworld TTS 1.5 lists 113 voice IDs in its schema, plus a temperature control for variability (see our 10 Inworld scripts). MiniMax Speech 2.8 HD pairs 17 preset voices with an emotion setting, pitch control and a language boost covering 38 options. Gemini 3.1 Flash TTS offers 30 prebuilt voices, a plain-language styling prompt and a two-speaker mode, and XTTS v2 speaks in a voice you supply as a sample, across 16 languages. Our text to speech ranking sorts them by job, and the full lineup lives under text to voice in the each::labs model catalog.

But try the 11 first. For most English narration and explainer audio, Bella, Emma, Michael or George will do the job.

Frequently Asked Questions

How many Kokoro 82M voices can I use on each::labs?

Eleven, all English. The voice field accepts af (the default), af_bella, af_sarah, af_nicole, af_sky, am_adam, am_michael, bf_emma, bf_isabella, bm_george and bm_lewis. That's the original v0.19 voice lineup, seven American and four British.

What does the af in a Kokoro voice name mean?

Two facts in two letters. The a is American English, and a b in that position would mean British English. The f marks a female voice, and m a male one. Plain af, with no name after it, is the default: a 50-50 blend of Bella and Sarah, according to Kokoro's v0.19 card.

Which Kokoro voice has the best grade?

On hexgrad's voice card, af_bella is the highest-graded voice of the 11, at A-. The next tier is af_nicole and bf_emma at B-, both trained on tens of hours of audio. The grades measure training data, so always confirm by ear with your own script.

Does the speed setting change the voice?

It changes pace, not the speaker. In our test, speed 1.25 made the same line about 14 to 16 percent shorter for af_bella and bm_george. Use it to fit timed slots or to calm a voice that rushes on long text, and pair it with paragraph-sized chunks.

Start with what you need.

Scale when you’re ready.