How the browser implementation works
The page loads kokoro-js and the ONNX Community model from public distribution services. WebGPU uses the fp32 model listed at 326 MB. Without WebGPU, the WebAssembly path requests the q8 model listed at 92.4 MB. The exact transfer can differ because browsers cache multiple supporting files. Generation begins only after the user selects Natural AI and starts a job.
The eight voices shown by VoiceAI Free
The curated list is Heart, Bella, Nicole, Fenrir, Michael, Emma, Fable, and George. The model card identifies their nationality and gender categories; the style and use-case labels on this site are editorial listening guides, not measured scores. Preview samples use original VoiceAI Free scripts and the same published model voices.
Limits worth knowing before a long project
The local editor limits a generation to 1,000 characters to reduce memory failures. Output is WAV. Current Natural AI mode is focused on English; multilingual device voices remain available for listening. The model repository lists Apache-2.0, but users must still hold rights to their input and check any voice, dependency, or distribution requirements that apply to the project.
Model source: ONNX Community Kokoro model card and file table, checked September 6, 2026.
Questions people ask
Is VoiceAI Free really free?
Yes. It has no account, subscription wall, or monthly credits. Device and browser limits still apply.
Does it upload my text?
No editor text is sent to a VoiceAI Free generation server or included in analytics. Device voices and third-party resource downloads have separate provider boundaries described in the privacy policy.
Can every output be used commercially?
No blanket promise is made. Check the text, voice, model, platform, and distribution rights for the specific project.