Kokoro-82M
hexgrad · Apache-2.0
The voice model, licensed under Apache 2.0. This component license is separate from VoiceCub’s service terms; free VoiceCub audio is for personal, non-commercial use.
Model & license ↗OPEN SOURCE, OPEN POSSIBILITIES
Built on good work.
VoiceCub is possible because people build things and share them. Here’s the open-source software behind the voices.
Kokoro ONNX weights
ONNX Community · Apache-2.0
The browser-compatible model conversion, downloaded directly from Hugging Face.
Model files ↗kokoro-js 1.2.1
hexgrad & Xenova · Apache-2.0
The JavaScript library that turns a script into audio.
Source code ↗Transformers.js
Hugging Face · Apache-2.0
Loads and runs the model locally, with browser caching.
Source code ↗ONNX Runtime Web
Microsoft · MIT
The WebGPU and WebAssembly inference runtime. Its runtime files are served from this site.
Source code ↗phonemizer.js 1.2.1
Xenova · Apache-2.0 wrapper; GPL-3.0 eSpeak NG inside
Converts written English into phonemes. The original, unmodified module is loaded separately.
Source & build instructions ↗eSpeak NG
eSpeak NG contributors · GPL-3.0
The speech engine compiled into phonemizer.js. Its GPL license applies to the embedded engine.
Engine source code ↗lamejs 1.2.1 / LAME
lamejs & LAME contributors · LGPL-3.0 (lamejs)
Encodes 128 kbps MP3s on your device. The original lame.min.js is kept as a separate, unmodified file and loaded after each voiceover to prepare the MP3 download.
Encoder source code ↗Licenses and replaceable components
Read the bundled license notices: kokoro-js, Transformers.js, ONNX Runtime, phonemizer.js, eSpeak NG GPL-3.0, lamejs notice, and LGPL-3.0.
The delivered modules can be inspected or replaced with compatible versions: phonemizer.js and lame.min.js. VoiceCub does not prevent modification or reverse engineering for debugging modifications to these components. Upstream source and build instructions are linked above. Learn more about LAME.
Voice samples
The 13 preview MP3s were generated with Kokoro in fp32 at normal speed, then encoded with FFmpeg / libmp3lame. The preview script and regeneration instructions are included with the project.
Made by
VoiceCub is operated by JI'AN WANGDONG TECHNOLOGY CO., LTD.
Room 806, No.801, Zongshanghui Building, No.9 Jinggangshan Avenue, Jizhou District, Ji'an, Jiangxi 343000, ChinaUnified Social Credit Code 91360802MADAP1JBXQ
[email protected]