Kokoro
Kokoro is a tiny open-weight text-to-speech model from the developer hexgrad, released as v1.0 on 27 January 2025 with 82 million parameters, 54 voices across eight languages, an Apache 2.0 licence and training data limited to permissive or synthetic audio. It has no emotion or style controls; you pick a voice and a speed and it reads. It returns in under a second, and it runs on a laptop CPU if you would rather not pay at all.
Anvisha Pai, Co-founder & CEO, Voyager
hexgrad · released
›What we asked for
Hi, I'm Kokoro, an open-weight voice model with eighty-two million parameters, small enough to run on a laptop. I don't do emotions or accents on request, I just read. Here is the same script every voice on this site reads, so you can compare us.
voice: af_heart, speed: 1
returned 16.1 seconds of wav
At a glance
- Price
- A fifth of the price of Eleven v3; our four scripts cost two cents.
- Voices
- Nineteen American English voices on the hosted endpoint, 54 across eight languages in the weights; we used af_heart.
- Open weights
- Apache 2.0, 82 million parameters, trained on permissive and synthetic audio only.
- Speed
- Every reading came back in about half a second.
- No expression
- No tags, emotions or cloning; the script is read as written.
Anvisha's take
We have not scored the four readings by ear; press play and judge the voice yourself. What we measured: Kokoro read every script in about half a second, twenty times faster than Eleven v3 and thirty times faster than Gemini, for two cents across all four. It also read the narration paragraph fastest of the four voice models we ran in September: 22.9 seconds, where the others took 24 to 26. It returns WAV, which the harness converts to mp3 for the page. With no emotion controls the emotional script is the honest test of a fixed-voice model; the names-and-numbers script tests its text normalisation, which is the usual weakness of small models.
Kokoro real output: the same 4 scripts every model gets
Every voice model on this site gets the same 4 scripts, so you can compare like with like. These are the readings Kokoro returned, one try each, nothing re-rolled or retouched. Generated on . The 4 timed readings took 0 to 1 seconds each and cost $0.02 in total.
Pacing, warmth and whether the pauses land where a voice actor would put them.
›What we asked for
Meet the Ember mug. It keeps your coffee at exactly the temperature you choose, for up to ninety minutes, and it tells your phone when it's ready. Warm from the first sip to the last. Ember. Coffee, on your terms.
voice: af_heart, speed: 1
returned 12.8 seconds of wav
Steady narration over a long paragraph, clean list rhythm and a natural close.
›What we asked for
A pour-over works in four steps. First, rinse the paper filter with hot water so it doesn't taste of paper. Second, add the grounds and pour just enough water to wet them, then wait thirty seconds while they bloom. Third, pour the rest of the water in slow circles. Fourth, wait about three minutes for it to drip through. That's it: a cup that tastes like the coffee, not the machine.
voice: af_heart, speed: 1
returned 22.9 seconds of wav
Range: excitement, then a hushed aside, from the words alone with no tags.
›What we asked for
We did it! We actually did it! I can't believe it's finally over. Okay. Okay, deep breath. Let's not tell anyone until Monday, alright? Just... let me have this one quiet night.
voice: af_heart, speed: 1
returned 11.4 seconds of wav
Reading numbers, times, currency, an email address and two hard names correctly.
›What we asked for
Your order ships on the 24th of September 2026 and arrives by 9:15 a.m. The total is $1,249.99, including VAT. Questions? Ask for Siobhan Nguyen or Ravi Parikh on 0800 555 0199, or email help@moda.app.
voice: af_heart, speed: 1
returned 21.1 seconds of wav
How the cost under each reading was worked out: The Kokoro list rate of $0.02 per 1,000 characters for the 213 characters sent.
Kokoro pricing
| Where | Unit | Price | Source |
|---|---|---|---|
| Self-hosted | 1,000 characters Free under Apache 2.0; runs on CPU. | $0.00 | hexgrad/Kokoro-82M on Hugging Face, read |
Is Kokoro free?
Kokoro is free to run yourself under Apache 2.0 and small enough to run on a CPU; several web demos host it free.
- •Open weights: Kokoro-82M on Hugging Face, Apache 2.0.
- •Hosted endpoints: priced per thousand characters.
Where to use Kokoro
- Open weights · Apache 2.0 weights and a pip package.
- Voyager · Runs inside Voyager, the agent for creative work, on Voyager credits on every paid plan, or with your own fal key.
Voyager
Run Kokoro inside Voyager
Voyager is an agent for creative work that runs every model in this reference, Kokoro included. Brief it, and it picks the model, makes the thing and hands back something you can edit. Create an account to get started.
What we noticed running Kokoro
- •Every reading returned in 0.5 to 1.0 seconds, by far the fastest voice model on this site.
- •22.9 seconds for the paragraph, the quickest of the four September readings; the others took 24 to 26. Gemini 3.8 Flash TTS, added on 22 October, read it in 22.6. See the explainer paragraph reading.
- •Returns WAV rather than mp3; the page plays an mp3 the harness made from it.
Kokoro limits and API parameters
| Limit | Value |
|---|---|
| Expression | None: no tags, emotion or cloning. |
| Languages on the endpoint | American English on this endpoint; British English and other languages are separate endpoints. API schema, read |
›Every setting, for developers
Kokoro American English text-to-speech
API schema, read
| Parameter | Type | Values | Default |
|---|---|---|---|
| prompt | string | The script Required. | |
| voice | enum | af_heart, af_alloy, af_aoede, af_bella, af_jessica, af_kore, af_nicole, af_nova, af_river, af_sarah, af_sky, am_adam, am_echo, am_eric, am_fenrir, am_liam, am_michael, am_onyx, am_puck, am_santa | af_heart |
| speed | number | 0.1 to 5 | 1 |
What Kokoro will not say
Apache 2.0 weights with no usage restrictions beyond the licence; the model cannot clone voices, which removes the main misuse.
Alternatives to Kokoro
Frequently asked questions
What is Kokoro TTS?
Kokoro is an 82-million-parameter open-weight text-to-speech model under Apache 2.0, released in January 2025 with 54 voices in eight languages. It runs on a CPU, and hosted endpoints charge per thousand characters.
Is Kokoro free?
Yes, to run yourself under Apache 2.0. Hosted endpoints charge per thousand characters.
Can Kokoro clone voices?
No. It has fixed voices and no cloning or emotion controls.
How this page is made
Prices, limits and settings are read on the dates shown, from the linked vendor pages or from the API schema we called. Every picture was generated by us with the request shown under it, and the original files are kept unedited. A generation that failed is shown as a failure. Model and vendor names are used to identify the products; no vendor imagery appears on this page.
