🦜 Kokoro MLX Voice Blender

Run Text-To-Speech with the MLX implementation (Mac M1-M4) of Kokoro to vastly improve processing speed. Use one voice or blend two voices by specifying a mixing ratio.

The app comes with a user-friendly gradio web interface.

Table of Contents

• Prerequisites
• Installation
• Usage
• Examples
• Gradio Web App
• Acknowledgment

Prerequisites

• Python >= 3.11
• HuggingFace Access Token

Installation

1. Clone this repo

git clone https://github.com/tsmdt/kokoro-MLX-blender.git

2. Change to project folder

cd kokoro-MLX-blender

3. Create a python env and activate it

python3 -m venv venv_kokoro
source venv_kokoro/bin/activate

4. Install kokoro-MLX-blender

pip install .

5. Download Kokoro MLX model using huggingface-cli

Run the following command from the main project folder (./kokoro-MLX-blender/)

huggingface-cli download --local-dir models/Kokoro-82M-bf16 mlx-community/Kokoro-82M-bf16

Ensure that the folder Kokoro-82M-bf16 (Hugging Face) with a voices subfolder and various .pt files (e.g. af_heart, af_alloy, etc.) now exists within the models folder. Your directory should look like this:

kokoro-MLX-blender
├── kb_mlx/
├── models/
│   └── Kokoro-82M-bf16/
│       ├── samples/
│       ├── voices/
│       ├── .gitattributes
│       ├── config.json
│       ├── DONATE.md
│       ├── kokoro-v1_0.safetensors
│       ├── README.md
│       ├── SAMPLES.md
│       └── VOICES.md
├── .gitignore
├── LICENSE
├── README.md
...

Note

You can use different versions of the KokoroMLX model as well. Download your preferred one from HuggingFace (cf. Installation step 5) and make sure that the downloaded Kokoro model folder exists within the models folder of kokoro-MLX-blender.

6. Check if everything works correctly

Run the following command in CLI to check if everything works.

kbx list

If you see a list of voice names kokoro-MLX-blender should work. If not please make sure that you downloaded the kokoro model in the previous step and placed it correctly in your models folder.

Usage

$ kbx run

 Usage: kbx run [OPTIONS]

 Run TTS with KokoroMLX for M1-M4. Use one voice or blend two voices.

╭─ Options ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ *  --text        -t                   TEXT       Input text(s) as string, single .txt or directory path [default: None] [required]   │
│    --voice1      -v1                  TEXT       Name of the first voice (without .pt) [default: af_heart]                           │
│    --voice2      -v2                  TEXT       Name of second voice (without .pt); if omitted, use only voice1 [default: None]     │
│    --mix-ratio   -m                   FLOAT      Blend weight for voice1 and voice2 (0.5 = 50% each) [default: 0.5]                  │
│    --speed       -s                   FLOAT      Speed multiplier (1.5 = 50% faster, 0.5 = 50% slower) [default: 1]                  │
│    --model-dir   -md                  DIRECTORY  Path to the local Kokoro model directory [default: ./models/Kokoro-82M-bf16]        │
│    --output-dir  -o                   TEXT       Directory where output audio file will be saved [default: ./output]                 │
│    --verbose          --no-verbose               Enable verbose output [default: verbose]                                            │
│    --help                                        Show this message and exit.                                                         │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

Examples

Run TTS with two blended voices (60% voice1 and 40% voice2)

kbx run -t "This is a test in blending the male American voice of Eric with the female American voice of Heart." -v1 am_eric -v2 af_heart -m 0.6

Gradio App

Launch the Gradio web app like this:

kbx app

Acknowledgment

OpenAI's o4-mini-high was used for creating the CLI app.

Citation

@software{kokoro_mlx_blender,
  author       = {Thomas Schmidt},
  title        = {Kokoro MLX Voice Blender},
  year         = {2025},
  url          = {https://github.com/tsmdt/kokoro-MLX-blender},
  note         = {Accessed: 2025-05-29}
}