Glenn Lyvers Artistry in Everything I Do

Tech · Text to speech

Kokoro voices: all 28 English voice samples

Kokoro has 28 English voices: 20 American (11 female, 9 male) and 8 British (4 female, 4 male). Every one of them is below, reading the same line at the same speed, so you can compare them side by side.

I made these samples in June 2026 while picking narrators for the audiobook of my novel Bloodsworth: Tidal Brood. Each voice says its own name, then one line from the book's premise, at speed 0.88. The grade next to each voice comes from the official Kokoro VOICES.md file. The notes are mine.

Which Kokoro voice is best?

By the model card's grades, af_heart (A) and af_bella (A-) are the strongest. For a British voice, bf_emma (B-) has the best grade. The card says the grades estimate the quality and quantity of each voice's training data, and that voices will sound better or worse to different people.

For long nonfiction I use bm_george, even though the card grades it only C. I locked it as the narration voice for my article site because it pronounces scripture names and uncommon words accurately. An earlier engine I tried, DramaBox, kept mispronouncing those, and Kokoro's phoneme-based reading fixed that. I run it 12% slower than default (speed 0.88), and it has now read more than 800 articles and sermons. One render ran 109 minutes in a single file.

So treat the grade as a starting point and listen. A low-graded voice can still be the right one for a particular kind of text.

All 28 English voices

Same script for every voice, with the voice's own name in the first sentence: "This is Heart. Strange sightings in the Chesapeake Bay drew them out to the island, where something ancient and patient was waiting." Speed 0.88, 24 kHz output encoded to 192 kbps MP3.

VoiceAccent, sexCard gradeTraining audioSampleNotes
af_heart American, female A not listed Download Top grade on the card, and the voice the model card's own code example uses.
af_alloy American, female C 10 to 100 minutes Download One of the two guest voices in my multi-voice podcast script.
af_aoede American, female C+ 1 to 10 hours Download
af_bella American, female A- 10 to 100 hours Download Second-best grade on the card, and one of three English voices listed with 10+ hours of training audio.
af_jessica American, female D 10 to 100 minutes Download
af_kore American, female C+ 1 to 10 hours Download Sample added November 2026 with the same script and speed.
af_nicole American, female B- 10 to 100 hours Download The card tags it with a headphones icon. Its sample here is the longest file of the set.
af_nova American, female C 10 to 100 minutes Download Sample added November 2026 with the same script and speed.
af_river American, female D 10 to 100 minutes Download
af_sarah American, female C+ 1 to 10 hours Download
af_sky American, female C- under 10 minutes Download Trained on less than 10 minutes of audio, per the card.
am_adam American, male F+ 1 to 10 hours Download Lowest grade of all 28 English voices.
am_echo American, male D 10 to 100 minutes Download The other guest voice in my podcast script.
am_eric American, male D 10 to 100 minutes Download
am_fenrir American, male C+ 1 to 10 hours Download On my shortlist of male narrators before I settled on bm_george.
am_liam American, male D 10 to 100 minutes Download
am_michael American, male C+ 1 to 10 hours Download On the same shortlist.
am_onyx American, male D 10 to 100 minutes Download On the same shortlist.
am_puck American, male C+ 1 to 10 hours Download
am_santa American, male D- under 10 minutes Download Trained on less than 10 minutes of audio, per the card.
bf_alice British, female D 10 to 100 minutes Download
bf_emma British, female B- 10 to 100 hours Download Highest-graded British voice on the card.
bf_isabella British, female C 10 to 100 minutes Download
bf_lily British, female D 10 to 100 minutes Download
bm_daniel British, male D 10 to 100 minutes Download
bm_fable British, male C 10 to 100 minutes Download
bm_george British, male C 10 to 100 minutes Download My locked voice for article narration: more than 800 articles and sermons so far.
bm_lewis British, male D+ 1 to 10 hours Download On my male shortlist. My notes paired it with bm_george as the more formal British option.

Grades and training durations: hexgrad/Kokoro-82M, VOICES.md. The card gives no training duration for af_heart.

What I learned narrating a whole novel

The Bloodsworth audiobook is 41 chapters and about 137,000 words, rendered twice so readers can pick a female or a male narrator. A few things I had to work out along the way:

Pacing comes from silence

Kokoro has no emotion or SSML controls. My render script can change the speed line by line, but slowing a voice down warps its pitch, like a record dragged on a turntable, so I left that switch off. So the book runs at one constant speed (0.8), and the drama comes from silence my script inserts: about two-thirds of a second between paragraphs and close to a second around chapter headings and scene breaks. A short punch line gets a little extra time to land.

Give it whole paragraphs

When I rendered sentence by sentence, short lines like "Patient." came out warbly. Feeding Kokoro a full paragraph at a time gave it enough context for natural phrasing. The model card says the same thing in its own terms: voices do best on 100 to 200 tokens, are weak on utterances under 10 to 20 tokens, and can rush past 400.

Fix words by ear

Some mistakes only show up when you listen. Kokoro read "breathed" wrong, so the script respells it "breethed" before rendering. Abbreviations get spelled out first ("Dr." becomes "Doctor"), and the time "1405" becomes "fourteen oh five". For the article narrator, "persecution" gets respelled "per-suh-kyoo-shun".

Clean the text first

I don't want dashes read as anything, so my scripts turn em dashes into commas, which Kokoro reads as a short pause. Watch for broken character encoding, too: when a dash got mangled into "â" plus stray bytes, Kokoro read the "â" out loud as "circumflex".

Match the language code to the voice

American voices use lang_code='a' and British voices use lang_code='b'. My podcast script mixes bm_george as host with af_alloy and am_echo as guests, so it keeps one pipeline per language code and sends each turn to the right one.

Kokoro voices FAQ

How many Kokoro voices are there?

Kokoro v1.0, released January 27, 2025, has 54 voices across 8 languages. 28 of them are English: 20 American and 8 British. The rest cover Japanese, Mandarin Chinese, Spanish, French, Hindi, Italian and Brazilian Portuguese. (Source: Kokoro-82M model card.)

What do af, am, bf and bm mean?

The first letter is the accent (a for American, b for British) and the second is the sex (f or m). The first letter also matches the lang_code you pass to the pipeline.

Can you blend Kokoro voices?

Yes. Pass several voice names separated by commas, such as voice="af_bella,af_heart", and the pipeline averages them with equal weight. You can see this in load_voice in kokoro/pipeline.py. I haven't used blends for narration myself; every file I've published uses a single voice.

What speed should I use?

The default is 1.0. I use 0.88 for articles and 0.8 for the novel. For the book I kept that one speed throughout and got the pacing from pauses instead.

Can I use Kokoro voices commercially?

The model weights are released under the Apache 2.0 license, and the model card says the model can be deployed anywhere from production environments to personal projects. Kokoro has 82 million parameters. My article narrator runs it on a CPU, and the novel was rendered on a GPU.

More from this site

Listen to the finished audiobook on the Bloodsworth: Tidal Brood page, see my tech mentoring and classes, or read more about me.