Question 1

What is Moss Voice?

Accepted Answer

Moss Voice is a browser-native text-to-speech workspace built on MOSS-TTS-Nano. You type a script, pick a voice (or clone your own), and the model generates speech entirely on-device via ONNX Runtime Web.

Question 2

Does Moss Voice send my text to a server?

Accepted Answer

No. Synthesis runs in the browser tab using the cached model. Only the model weights themselves pass through the edge cache on the first download. Script text, generated audio, and any voice you clone never leave the tab.

Question 3

How does voice cloning work?

Accepted Answer

Drop or record a 10-30 second clip. The browser resamples it to 48 kHz stereo, runs the MOSS-Audio-Tokenizer-Nano encoder ONNX model on-device to extract prompt audio codes, and stores those codes in OPFS as a new voice you can use immediately.

Question 4

What languages does Moss Voice support?

Accepted Answer

Any voice can read text in 19 languages: Chinese, English, German, Spanish, French, Japanese, Italian, Hungarian, Korean, Russian, Persian, Arabic, Polish, Portuguese, Czech, Danish, Swedish, Greek, and Turkish.

Question 5

How big is the model download?

Accepted Answer

About 745 MB on first visit (TTS network + audio tokenizer + encoder for voice cloning). After that the bytes live in your browser's OPFS cache and subsequent visits skip the download.

Question 6

Is Moss Voice free and open source?

Accepted Answer

Yes. The web app is MIT-licensed; the underlying MOSS-TTS-Nano and MOSS-Audio-Tokenizer-Nano models are Apache-2.0 from the OpenMOSS team.

Type a script. Pick a voice. Hear it in seconds.