# Kokoro TTS: what it is, how it sounds and how to run it

> Kokoro is an open text to speech model with 82 million parameters and an Apache licence. Its voices, languages, limits, and three ways to run it offline.

Canonical: https://bitfaceless.com/blog/kokoro-tts.md

[Production](https://bitfaceless.com/blog/category/production)

Nemo · November 17, 2026 · 7 min read

Kokoro TTS is an open text to speech model that turns written text into spoken audio on your own computer. It has 82 million parameters, which is small for a voice model, its weights are Apache licensed, and version 1.0 ships 54 voices across 8 languages. You can run it with a few lines of Python, in a web browser, or inside an app that has it built in.

## What is Kokoro TTS?

Kokoro is published by its author on Hugging Face as [Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M), with the code in [the kokoro repository on GitHub](https://github.com/hexgrad/kokoro). The facts below come from those two pages.

- **Size.** 82 million parameters. The model card describes it as an open-weight model in that size class.
- **Licence.** Apache 2.0 for the weights and the code. The model card says it can be used in production and personal projects.
- **Versions.** Version 0.19 came out on December 25, 2024 with 1 language and 10 voices. Version 1.0 came out on January 27, 2025 with 8 languages and 54 voices.
- **Training data.** A few hundred hours of audio that is permissively licensed or not under copyright, according to the card.
- **Output.** Audio at 24,000 Hz.

One warning from the model card is worth repeating: websites with “kokoro” in their domain name are not owned by the author or affiliated with the model. The official sources are the Hugging Face page and the GitHub repository.

## Which languages and voices does Kokoro have?

The [official voice list](https://huggingface.co/hexgrad/Kokoro-82M/blob/main/VOICES.md) names every voice and grades each one by the quality and amount of its training data.

| Language | Code | Voices |
|---|---|---|
| American English | a | 20 (11 female, 9 male) |
| British English | b | 8 (4 female, 4 male) |
| Mandarin Chinese | z | 8 |
| Japanese | j | 5 |
| Hindi | h | 4 |
| Spanish | e | 3 |
| Brazilian Portuguese | p | 3 |
| Italian | i | 2 |
| French | f | 1 |

More than half of the voices are English, and English is where Kokoro is strongest. The voice list says support outside English can be thin because of limited training data, weak conversion from text to sounds, or both.

## What does Kokoro sound like, and where does it fall short?

In English, Kokoro reads clearly and steadily, like a calm narrator. For explainers, tips, lists and tutorials, that is usually enough. Try a few voices and judge for yourself, because voice quality is partly taste. The voice list says so too.

The limits are documented, which is rare and useful:

- **Very short lines are weaker.** The voice list says quality drops on inputs under about 10 to 20 tokens. A one word line can sound off.
- **Very long passages rush.** Past about 400 tokens, voices tend to speed up. The fix is to split the text into chunks or lower the speed setting.
- **There is a comfortable range.** Most voices do best at around 100 to 200 tokens per passage.
- **Fixed voices.** You choose from the list. The model card does not describe a way to clone your own voice.
- **Less acting.** The top cloud voices handle emotion, pauses and emphasis with more range. Kokoro reads well. It does not perform.

## How does Kokoro compare with cloud voices?

Prices here were checked in November 2026 on each provider’s own page.

| Voice | What it makes | Where it runs | How it charges | Who it suits |
|---|---|---|---|---|
| Kokoro | Speech from text with 54 fixed voices | Your own computer or browser | Free, Apache licence | Offline English narration at any volume |
| [ElevenLabs](https://elevenlabs.io/pricing) | Expressive speech, voice cloning, sound effects, music | Cloud | Monthly credits: 10,000 free, then plans from $6 a month | Storytelling where delivery carries the video |
| [OpenAI speech API](https://developers.openai.com/api/docs/pricing) | Speech from text | Cloud, through an API | $15 per million characters for tts-1 | Developers and apps |

The cloud voices are better actors and cover more languages. Kokoro has no meter and does not send your script anywhere. If you are weighing the paid options, the post on [ElevenLabs alternatives](https://bitfaceless.com/blog/elevenlabs-alternative) compares them on price.

## How do you run Kokoro?

There are three routes, from most hands-on to least.

1. **With Python.** The repository’s instructions are to install the `kokoro` and `soundfile` packages with pip, plus the `espeak-ng` system package. Then you create a pipeline with a language code (`a` for American English, `b` for British), pass it your text and a voice name such as `af_heart`, and save the audio it returns. On an Apple silicon Mac, the README says to set `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the graphics chip. Japanese and Mandarin need extra packages.
2. **In a browser or Node.** The [kokoro-js package](https://github.com/hexgrad/kokoro/tree/main/kokoro.js) runs the model fully in the browser, on WebAssembly or WebGPU, or on the processor in Node. This is the route for web developers.
3. **Inside an app.** Some desktop apps include Kokoro, so you never touch a terminal. You pick a voice from a menu and the app handles the model, the text splitting and the timing.

Route 1 gives you full control and takes some comfort with a terminal. Route 3 gives you a narrator and nothing to maintain. In every case the model runs on your machine, so once it is downloaded you can generate speech with no connection.

## Kokoro inside Bitfaceless

Bitfaceless is a desktop app that turns a script or an idea into a short narrated vertical video hosted by a pixel avatar. Its local voices are built on Kokoro. They are English, US and UK, male and female, and they run offline on your computer with no per character cost.

The app does the work around the voice: it splits the script into lines, times word by word captions from the narration, and moves the avatar’s mouth with the audio. The Kokoro voice ships inside the app, so it works offline from the first video; a full-precision version is an optional download in Settings.

The limits follow from the above. The local voices are English only, there is no voice cloning, and the app is for making videos, so it is not a general voice tool. If you want a more expressive voice, you can connect ElevenLabs with your own API key, billed by ElevenLabs. The app runs on Mac and Windows (Apple silicon on macOS 14 or later, or a 64-bit PC on Windows 10 or 11). It is paid once with a 7-day refund: [see the price](https://bitfaceless.com/#pricing) or [download Bitfaceless](https://bitfaceless.com/download).

A local voice reads exactly what you give it, so the writing has to work when heard: see [how to write a script for a faceless video](https://bitfaceless.com/blog/how-to-write-a-faceless-video-script). For the wider question of local against cloud models, read [local AI vs cloud AI for making videos](https://bitfaceless.com/blog/local-ai-vs-cloud-ai-video).

## Frequently asked questions

### Is Kokoro TTS free?

Yes. The model weights and the code are published under the Apache 2.0 licence, and running the model on your own computer has no usage fee. Hosted services that offer Kokoro through an API set their own prices.

### Can I use Kokoro TTS commercially?

The model card says the Apache-licensed weights can be used in production as well as personal projects. That covers narration for videos you earn from. Read the licence itself if you plan to ship the model inside a product.

### Does Kokoro support voice cloning?

The official model card and voice list describe a fixed set of 54 voices and no cloning feature. If you need a copy of a specific voice, you need a different tool, and the right to use that voice.

### Does Kokoro run offline?

Yes. After the model files are on your disk, speech is generated on your own machine. The kokoro-js package even runs it inside a browser tab without sending text to a server.

### What languages does Kokoro TTS support?

Version 1.0 lists American and British English, Japanese, Mandarin Chinese, Spanish, French, Hindi, Italian and Brazilian Portuguese. English has 28 of the 54 voices and the best quality. Some languages have only one to three voices.

### Is Kokoro as good as ElevenLabs?

Not for expressive delivery. ElevenLabs voices have more emotional range, more languages and cloning. Kokoro is good enough for clear narration, costs nothing per character and works offline, which is the trade many faceless channels want.

Your avatar narrates. Your video's done.

Bitfaceless turns a script or an idea into a narrated video hosted by a pixel avatar, on your own computer. $39 once, 7-day refund.

 [See how Bitfaceless works](https://bitfaceless.com/#how-it-works)

**About this article.** Written by Nemo, who builds Bitfaceless. Facts about other tools and platforms were checked on the date above and can change. Spotted something wrong or out of date? [Tell us](mailto:support@bitfaceless.com).

## More in Production

[Production](https://bitfaceless.com/blog/category/production) Nov 17, 2026 · 7 min

## [How to add captions to a video: 4 ways compared](https://bitfaceless.com/blog/how-to-add-captions-to-a-video)

Four ways to add captions to a video: platform auto captions, an editor's auto-caption tool, an SRT file, and captions timed from the narration. Steps for each.

[Production](https://bitfaceless.com/blog/category/production) Nov 14, 2026 · 4 min

## [Local AI vs cloud AI for making videos](https://bitfaceless.com/blog/local-ai-vs-cloud-ai-video)

What changes when the AI that writes, voices and illustrates your videos runs on your own computer instead of a cloud service: cost, privacy, speed and quality.

[Production](https://bitfaceless.com/blog/category/production) Nov 15, 2026 · 4 min

## [Why give a faceless channel an avatar host](https://bitfaceless.com/blog/avatar-host-for-faceless-videos)

Faceless channels blur together because nobody hosts them. How an avatar host makes a channel recognizable, what it should do on screen, and how to design one.
