Paradee-8M v1.0

Paradee-8M v1.0

Speaks English text in a single voice with an 8M-parameter model distilled from Kokoro-82M. The int8 ONNX file is about 9 MB and runs on a CPU. Suits developers who need tiny, fast speech output with no GPU.

Made by sahilmahendrakar 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: Paradee-8M turns English text into speech with just 8.07 million parameters, and it learned its voice from the larger Kokoro-82M model. It speaks one voice, Kokoro's af_heart. The int8 version is a single ONNX file of about 9 MB, compared with 325 MB for Kokoro, and it runs on a CPU without a GPU.

Who it suits: Developers who need small, quick speech output in an app, a browser or an edge device, under the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Software | App developers | Add spoken output with a 9 MB model | Read a message aloud
Web | Front-end developers | Run speech in the browser through ONNX tools | Speak a page summary
Embedded and edge | Hardware teams | Run voice on a CPU-only device | Speak status messages
Research | Speech researchers | Study distilling a large voice model into a tiny one | Compare Paradee with its teacher

Skip it if you need several voices or languages other than English, since it has one voice and one language.

Where to use it: It runs as an ONNX file or through the author's Python package.

• onnxruntime on a CPU
• The paradee Python package, with a command-line tool
• Browsers, through the kokoro-js tokenizer files the repository includes
• Windows, macOS or Linux

Use case | Who | What they get back
On-device speech | Developers | A 24 kHz waveform from text
Command-line speech | Scripters | A wav file from one command
Browser speech | Web developers | Audio from phoneme input

How to set it up: You need Python and the model files.

• Python with pip
• Disk space for the 9 MB int8 file, or the larger fp32 file
• The paradee Python package from the author's GitHub project

1. Install the Hugging Face hub tools with pip.
2. Download the repository with the hf command.
3. Install the paradee package from the author's GitHub project.
4. Run the command line or the Python code below.

pip install -U huggingface_hub
hf download sahilmahendrakar/Paradee-8M-v1.0

How to use it: Run the command-line tool with a sentence and an output file name, or call the model from Python and write the wav file yourself.

python -m paradee "Paradee is a small voice that runs anywhere." -o hello.wav

from paradee import Paradee, SAMPLE_RATE
import soundfile as sf

tts = Paradee()
sf.write("hello.wav", tts("Paradee is a small voice that runs anywhere."), SAMPLE_RATE)

Tool | What it does
onnx/paradee_int8.onnx | The whole model in one int8 graph, which the card says to use
onnx/paradee.onnx | The same graph in fp32, at 37 MB, which sounds the same
config.json | Holds the phoneme vocabulary and the sample rate
tokenizer.json | The same vocabulary for transformers.js and kokoro-js
python -m paradee | Writes a wav file from a sentence

Example: You run the command with a short sentence and the output name hello.wav. A file with that sentence spoken in the af_heart voice appears in the current folder.

For newcomers and experts: If you are new to speech models, run the one-line command and listen to the wav file. It shows that a tiny file can talk.

If you know the field, the ONNX graph takes phoneme ids and a speed value. Each id list has a pad token at both ends and at most 512 entries. The output is a 24 kHz waveform. Phonemes must be written the way Kokoro's misaki library writes them, since that is all the model saw in training. In browsers, a conversion step from eSpeak spelling is needed.

Turn text into misaki phonemes
Pass phoneme ids and speed to ONNX
Run the int8 graph on a CPU
Get a 24 kHz waveform
Save or play the audio

Good to know: The card reports a UTMOS naturalness score of 4.41 for Paradee and 4.52 for Kokoro, with the same word error rate of 5.7 percent. The card measures the released ONNX file at 18 times real-time speed on one CPU thread.

Title: File size in MB
Kokoro-82M: 325 MB
Piper: 63.2 MB
KittenTTS nano: 56.8 MB
Kokoro-7M-Distill: 30.1 MB
Paradee: 8.45 MB

• License: Apache 2.0
• Cost to run: none named, apart from your own hardware
• Data it reaches: the text you give it and the model files on your computer, and nothing else the card names
• Not for: other voices or languages, since the model speaks one English voice

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: onnx
• Downloads in the last 30 days: 4,374
• Likes: 70
• License: apache-2.0
• Languages: en
• Based on: hexgrad/Kokoro-82M
• Quantized versions: 3
• Tags: onnx, paradee, text-to-speech, distillation, kokoro

Links

• Model: https://huggingface.co/sahilmahendrakar/Paradee-8M-v1.0
• Files: https://huggingface.co/sahilmahendrakar/Paradee-8M-v1.0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sahilmahendrakar/Paradee-8M-v1.0

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