Kokoro Voices: Full List, Grades, and Accents

Kokoro v1.0 ships 54 preset voices across eight languages and nine accent groups, and each voice carries an upstream grade that estimates the quality and quantity of the audio data behind it. The voice ID tells you the group at a glance: af_ is American English female, bm_ is British English male, jf_ is Japanese female. Use the grade to build a shortlist, then audition your own text before you commit, because a letter grade is a data estimate and not a verdict on how a voice fits your script.

This guide lists the complete Kokoro v1.0 catalogue by accent, explains what the grades actually measure, and gives a repeatable way to choose. The voice IDs, genders, training-duration bands, and grades below come from the official Hexgrad voice catalogue, checked on 2026-10-14. Kokoro is a preset-voice model, so none of these voices is cloned from a recording. If the job needs a specific person's voice, that is a different kind of tool.

What the Kokoro voice list actually contains

Every Kokoro voice ID is built from three parts: a language letter, a gender letter, and a name. The first letter is the language or accent code (a for American English, b for British English, j for Japanese, z for Mandarin Chinese, e for Spanish, f for French, h for Hindi, i for Italian, p for Brazilian Portuguese). The second letter is f or m for the target gender of the reference data. Everything after the underscore is the voice name, such as heart, bella, emma, or siwis.

That scheme makes the list easier to read than a random catalogue. You can filter by accent without opening every sample, and you can tell whether a group is deep or thin. English is by far the richest group, with 20 American voices and 8 British voices. French has exactly one voice. Several other languages were trained on only a few hours of audio, so support can be weak. The model's own catalogue warns that non-English output may be absent or thin because of weak grapheme-to-phoneme processing or a lack of training data.

How the Kokoro voice grades work

Each voice on the official catalogue carries a Target Quality grade, a Training Duration band, and an Overall Grade. The overall grade is meant to estimate both the quality and the quantity of the voice's training data, and the catalogue states plainly that these are estimates, that voices will sound better or worse to different listeners, and that the grade is not a subjective ranking.

• Target Quality rates the reference voice itself. Clean reference audio with matching text labels scores higher. Compression, artifacts, sample-rate problems, or a mismatch between the labels and the audio lower it.
• Training Duration rates how much audio the voice saw. The bands are HH hours (10 to 100 hours), H hours (1 to 10 hours), MM minutes (10 to 100 minutes), and M minutes (1 to 10 minutes). More data tends to mean a more stable voice.
• Overall Grade combines the two. A high grade usually means both clean audio and a long training set, while a low grade points to a small dataset, a noisy one, or both.

The Overall Grade uses a school-style scale with plus and minus steps. In the v1.0 catalogue the range runs from A (af_heart) and A- (af_bella) down to F+ (am_adam). That spread is the single most useful signal in the list, because it separates the handful of heavily trained voices from the long tail of lightly trained ones.

The full Kokoro v1.0 voice list by accent

American English voices (20)

American English is the largest and strongest group. If your script is General American, start here. Two voices stand out: af_heart sits at the top of the list with an Overall Grade of A, and af_bella follows at A- with an A target-quality rating and an HH-hours training band. Most of the male voices grade lower, and am_adam, which many people try first because the name is familiar, is at the bottom of the table with F+.

British English voices (8)

The British group is smaller but has a clear standout. bf_emma is the highest graded British voice at B-, helped by an HH-hours training band, and it is a common default for a classic British read. The rest cluster around C and D. If you need a British narrator, shortlist bf_emma first, then test bm_fable or bm_george for a male read.

Japanese and Mandarin Chinese voices (13)

The Japanese group lists five voices and the Chinese group lists eight. Japanese is the stronger of the two in this table, with all five voices at B target quality and grades between C+ and C-. The Mandarin voices all grade D, which reflects smaller training sets rather than any problem with the voices themselves. Test these on your exact text before trusting them with a full project, especially names and unusual readings.

Spanish, French, Hindi, Italian, and Brazilian Portuguese voices (13)

This group is where the catalogue thins out. Spanish has three voices and Brazilian Portuguese has three, but neither group lists quality or duration grades on the official table, so there is less to compare. French is represented by a single voice, ff_siwis, which grades B- on under 11 hours of data. Hindi and Italian each have a handful of B-quality voices graded C. If your target is one of these languages, plan to audition rather than assume.

A checklist for choosing a Kokoro voice

The grade narrows the field, but it does not tell you which voice suits your script. Work through these steps in order so the decision is repeatable and defensible, especially on client or branded work.

1. Pick the accent and language first. Filter the list to the group that matches your audience, then ignore every voice outside it.
2. Shortlist by grade. Start with the A, A-, and B- entries in your group, then add one or two C voices only if the higher graded options sound wrong for the tone.
3. Audition with your own text. Generate the same paragraph, at the same speed, for each candidate. A three-voice shortlist is enough to make a decision.
4. Score errors before taste. Note missing words, mispronounced names, and awkward pauses before you rank how pleasant each voice sounds.
5. Retest the winner on a representative long passage. Shortlist success does not prove a voice stays stable across a whole chapter or course.

For a ready-made worksheet, use the Kokoro voice audition scorecard; it has an original test passage and a CSV you can fill in yourself. If a candidate mispronounces your terms, treat it as a text problem first and check the pronunciation test suite before rejecting the voice.

Accents, languages, and limits to expect

Voice IDs tell you the intended accent, not the only accent the model can produce. The catalogue separates American (a) and British (b) English data, and a voice trained on one will not switch to the other. On accents beyond English, remember that some languages are covered by only a small handful of voices or, in the case of French, a single one. A low grade or a thin group is a caution to test more, not an automatic rejection.

The catalogue also describes a working range the voices were trained on: a goldilocks zone of roughly 100 to 200 tokens out of about 500 possible. Very short utterances can sound weak, and very long utterances can rush. The practical fixes are to merge short lines, chunk long passages into smaller units, or adjust the speed parameter. Kokoro does not expose cloning or natural-language voice design, so if those are requirements, compare model families on the local TTS model list instead of forcing Kokoro to do a job it was not built for.

Murmur is one way to hear and use these presets without managing a Python environment. It is a local voice studio for Apple Silicon Macs running macOS 15 or later, and it bundles Kokoro with an in-app preset picker that shows friendly names and accents, so you can audition and export on your own machine. When a voice misreads a specific term, the fix is often punctuation or spacing rather than a new model, and the text normalization guide covers those cases.

Be clear about the fit. Murmur is macOS-only, with no Windows, iPhone, Android, or browser version, and it does not do live streaming text to speech. Kokoro presets cannot clone a speaker; for that you would add a separate cloning model. Murmur costs $49 one-time with a free plan and a 3-day Pro trial, and a 7-day refund policy, and it runs speech generation locally, while activation, license validation, updates, and model downloads still use the network. If you only write for Windows, or you need cloud collaboration, a different tool is the honest answer.

Frequently asked questions