You can deploy this model and start earning money today!

gushilabs/
gushilabs-voices-for-kokoro-v1

Name: gushilabs-voices-for-kokoro-v1
Availability: InStock

Views

No views yet

gushilabs-voices-for-kokoro-v1

⚠️ Disclaimer: This is an unofficial, community-trained custom voice collection and encoder release. It is designed to be compatible with the Kokoro V1.0 release, but it is not created, endorsed, or maintained by the original author of Kokoro TTS. Please do not direct support questions regarding these specific voice packs to the official Kokoro repository.

This repository contains a custom-trained voicepack encoder and a collection of multi-speaker voices designed to work seamlessly with Kokoro TTS V1.0.

🎧 Speech Demos

Voice: Tony (Male)

Voice: Vivien (Female)

💻 Usage

python

# Uses the kokoro library as underlying framework
from kokoro import KPipeline

pipeline = KPipeline(lang_code="a", repo_id="hexgrad/Kokoro-82M")
voicepack = pipeline.load_voice("voices/af_vivien.pt")

text = "Hello world! This is Vivien. We skipped the diffuser denoiser, kept the style encoders, and now we are officially up and running."
audio = pipeline(text, voice=voicepack)

🛠️ Extracting Custom Voicepacks

If you would like to extract your own voicepacks (.pt files) from audio using our fine-tuned checkpoints, you can do so by running both style encoders (acoustic and prosodic) on representative audio files and averaging the resulting style vectors.

To run the extraction pipeline:

Clone our GitHub repository: train-kokoro-encoder-styletts2 and follow the installation instructions in its README.

Run the extract_voicepack.py script, pointing it to one of the .pth checkpoints from this Hugging Face repository and a folder of your target speaker's .wav files:

bash

python scripts/extract_voicepack.py \
    --model /path/to/epoch_2nd_00009.pth \
    --audio-dir /path/to/speaker_audio_folder \
    --output voices/my_custom_voicepack.pt

📊 Model Details

Property	Value
Architecture	StyleTTS2 (Kokoro-compatible)
Base Model	Kokoro V1.0
Language	English (en)
Training Epochs	[20]
Sample Rate	24 kHz

Trained and maintained by Gushi Labs.

gushilabs/gushilabs-voices-for-kokoro-v1