Kokoro-82M

Kokoro-82M

Turns text into speech with a small open-weight model of 82 million parameters, with 54 voices in version 1.0. For developers who want fast, low-cost speech in their apps.

Made by hexgrad on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Kokoro is a text-to-speech model with 82 million parameters and Apache-licensed weights. The card says it matches the quality of larger models while running faster and costing less. Version 1.0 lists 8 languages and 54 voices, and the model is built on StyleTTS2-LJSpeech.

Who it suits: It suits developers and makers who want natural speech from a small model they can run themselves or buy through a hosted API.

Industry | Who uses it | How they use it | Example
Software | App developers | Add spoken replies to an app | A voice that reads out chat answers
Accessibility | Product teams | Read on-screen text aloud | A reader for articles and documents
Media | Content creators | Narrate short videos and explainers | A voice-over made from a script
Education | Course makers | Turn lesson text into audio | Audio versions of study notes

If you want to clone a specific person's voice, you can skip it, since the card says its training used no custom voice clones.

Where to use it:

• A Python script or a notebook such as Google Colab
• Windows, macOS or Linux machines with Python
• A demo Space on Hugging Face, for listening before you install anything
• Hosted APIs, where the card says the model has been deployed commercially

Use case | Who | What they get back
Narration | Creators | WAV audio of your script in a chosen voice
App voice output | Developers | Speech generated on demand, chunk by chunk
Trying voices | Hobbyists | Sample audio in any of the listed voices

How to set it up: You need these first:

• Python and pip
• The espeak-ng speech tool, which the example installs with apt-get
• The kokoro package at version 0.9.2 or newer, plus soundfile

1. Install kokoro and soundfile.
2. Install espeak-ng.
3. Create a pipeline for your language code, then call it with your text and a voice name.

!pip install -q kokoro>=0.9.2 soundfile
!apt-get -qq -y install espeak-ng > /dev/null 2>&1

If you only want the weights, this downloads them.

pip install -U huggingface_hub
hf download hexgrad/Kokoro-82M

How to use it: Set a variable called text to the words you want spoken. Then create the pipeline and loop over what it yields, writing each chunk of audio to a WAV file at 24000 Hz.

from kokoro import KPipeline
import soundfile as sf
import torch
pipeline = KPipeline(lang_code='a')
generator = pipeline(text, voice='af_heart')
for i, (gs, ps, audio) in enumerate(generator):
    print(i, gs, ps)
    sf.write(f'{i}.wav', audio, 24000)

Tool | What it does
KPipeline | Builds the speech pipeline for a language code
voice | Picks the voice, such as af_heart
sf.write | Saves each audio chunk as a WAV file at 24000 Hz
misaki | The G2P library that the kokoro package uses to turn text into phonemes

Example: You set text to a short paragraph and run the loop. Kokoro splits the speech into chunks, prints each one, and saves files named 0.wav, 1.wav and so on that you can play in any audio player.

For newcomers and experts: It is a small voice engine that reads text aloud. A good first step is the demo Space, then the short Python loop above with one sentence.

The model is decoder-only, built on StyleTTS 2 and ISTFTNet, with no diffusion and no released encoder. It was trained on a few hundred hours of audio, using about 1000 hours of A100 80GB time.

Your text goes into the pipeline
misaki turns it into phonemes
The model generates the speech audio
Chunks are saved as WAV files

Good to know: Training was cheap by the card's own numbers.

Title: Training cost in USD
v0.19: 400 USD
v1.0: 600 USD
Total: 1000 USD

• Fake sites: the card warns that any website with kokoro in its domain name is not run by this model page or its author, and may be a scam
• Training data: only permissive or non-copyrighted audio, including public domain audio, Apache or MIT-licensed audio and synthetic audio from closed TTS models
• Other versions: 119 quantized and 71 fine-tuned versions made by other people are listed
• License: Apache 2.0
• Cost to run: none named for running it yourself, and the card quotes under $1 per million input characters on hosted APIs
• Data it reaches: the text you give it and the WAV files it writes

On Hugging Face, 13 November 2026

• Task: text to speech
• Downloads in the last 30 days: 12,580,963
• Likes: 7,403
• License: apache-2.0
• Languages: en
• Based on: yl4579/StyleTTS2-LJSpeech
• Quantized versions: 121
• Fine-tunes: 74
• Tags: text-to-speech

Links

• Model: https://huggingface.co/hexgrad/Kokoro-82M
• Files: https://huggingface.co/hexgrad/Kokoro-82M/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hexgrad/Kokoro-82M

Then follow the usage in its model card. Nothing in this kit runs by itself.