Kokoro-82M-v1.1-zh

Kokoro-82M-v1.1-zh

Speaks English and Chinese text aloud with an 82M-parameter text-to-speech model. A short training run on Kokoro that adds 100 Chinese speakers, for developers who want small local voices.

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-82M-v1.1-zh is a small text-to-speech model from hexgrad with 82 million parameters, covering English and Chinese. A brief extra round of training on the base Kokoro-82M gave it 100 Chinese voices, using data from a professional dataset company, plus a few hours of crowdsourced synthetic English. The release lists 103 voices across two languages.

Who it suits: Developers who add spoken output to apps or tools in English or Chinese, and people who want to try the new voices and report back. The model carries the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Software | App developers | Add spoken replies to apps in English and Chinese | A reader that speaks an article aloud
Media | Creators | Draft narration with a small local voice model | A Chinese voice-over draft
Research | Speech tinkerers | Test the new voices and tokenization and send feedback | Comparing its voices with the v1.0 release

Anyone who needs the larger voice set of the earlier release can skip it, since this one drops many voices.

Where to use it:

• Python with the kokoro package
• Google Colab, where the card says you can run its sample cell
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
English speech | App developers | Audio from English text
Chinese speech | App developers | Audio from Chinese text
Voice trials | Researchers | Samples of the new voices to judge

How to set it up:

• Python
• The kokoro and misaki packages at version 0.8.2 or newer, plus soundfile
• espeak-ng, which the card installs with apt-get

1. Install Python.
2. Install the packages with the first line of the first block below.
3. Install espeak-ng with the second line.
4. Optionally copy the model files with Hugging Face's two commands in the second block.

!pip install -q kokoro>=0.8.2 "misaki[zh]>=0.8.2" soundfile
!apt-get -qq -y install espeak-ng > /dev/null 2>&1

pip install -U huggingface_hub
hf download hexgrad/Kokoro-82M-v1.1-zh

How to use it: Usage matches the base Kokoro-82M, except that you pass the repo_id of this model when you build a KModel or a KPipeline. The repository's samples folder holds two scripts that make an English and a Chinese sample as audio files.

Tool | What it does
KPipeline | Builds the speech pipeline, taking this model's repo_id
KModel | Loads the model, taking this model's repo_id
make_en.py | Sample script that makes the English audio
make_zh.py | Sample script that makes the Chinese audio

Example: You run the card's notebook cell. It makes an English sample and a Chinese sample as WAV files, then plays them back at a 24000 sample rate.

For newcomers and experts: The first small thing to try is the card's Colab cell, which produces both samples without any setup of your own. Listen to them before you write code.

The architecture is StyleTTS 2 with ISTFTNet, decoder only, with no diffusion and no encoder released. It has the same 82 million parameters as the base model. The release trained on more than 100 hours of data and is meant to gather feedback on voices and tokenization before the series moves on.

Install kokoro and espeak-ng
Build a pipeline with this repo_id
Pass in English or Chinese text
The model speaks it
Save or play the audio

Title: Releases in the card
Model | Training data | Languages and voices
v1.1-zh | Over 100 hours | 2 languages, 103 voices
v1.0 | A few hundred hours | 8 languages, 54 voices
v0.19 | Under 100 hours | 1 language, 10 voices

Title: A100 80GB GPU hours per release
v0.19: 500
v1.0: 500
v1.1-zh: 120

Good to know:

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the download fetches files from Hugging Face, and the card's sample cell fetches its scripts from the repository
• Not for: anyone who needs every voice of the earlier release, because the card says this one is not a strict upgrade
• The Chinese training data was granted by a professional dataset company, and the English part is crowdsourced synthetic data

On Hugging Face, 13 November 2026

• Task: text to speech
• Downloads in the last 30 days: 76,178
• Likes: 247
• License: apache-2.0
• Based on: hexgrad/Kokoro-82M
• Quantized versions: 5
• Fine-tunes: 1
• Tags: text-to-speech

Links

• Model: https://huggingface.co/hexgrad/Kokoro-82M-v1.1-zh
• Files: https://huggingface.co/hexgrad/Kokoro-82M-v1.1-zh/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hexgrad/Kokoro-82M-v1.1-zh

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