Load the model
Select Load model. The ONNX graphs and the audio codec are fetched once and cached by your browser.
Text to Speech Online Free
Type a sentence and MOSS-TTS-Nano reads it back inside your browser. The model is a 0.1B multilingual speech model that runs on WebAssembly, so nothing is uploaded and no server is involved.
Your audio appears here
Model files
The speech model and audio codec come from Hugging Face, not from 73hi.
Weights live at MOSS-TTS-Nano-100M-ONNX and MOSS-Audio-Tokenizer-Nano-ONNX, both released under Apache-2.0.
Three simple steps
How to Turn Text Into Speech Online
Generate speech on your own machine without an account or a server queue.
Type and choose a voice
Enter your sentence, pick a built-in voice, or upload a short reference clip to clone a voice.
Play and save
Select Speak, listen in the player, then download the audio as a WAV file.
How it works
A 0.1B Speech Model, Entirely on Your Device
MOSS-TTS-Nano pairs a small autoregressive language model with a neural audio codec. Everything runs on the WebAssembly backend of ONNX Runtime Web.
| Stage | What it does | Where it runs |
|---|---|---|
| Tokenizer | Turns your text into the token ids the language model expects | Your browser |
| Prefill | Reads the text and voice prompt in one pass and produces the conditioning hidden states | Your browser |
| Frame decoding | Predicts one frame of audio codes at a time across 16 codebooks | Your browser |
| Audio codec | Decodes the audio codes back into a 48 kHz waveform | Your browser |
Free AI text to speech
Private, Multilingual and CPU Friendly
Everything needed to read text aloud without handing it to a third party.
Nothing is uploaded
Text and audio stay on your device. There is no request that carries what you typed.
No GPU required
The model is small enough to run on CPU through WebAssembly at usable speed.
Multilingual
MOSS-TTS-Nano supports many languages, including Chinese and English.
No quota
No signup, no daily limit and no watermark on the audio you generate.
Questions and answers
Text to Speech FAQ
Is my text uploaded?
No. The model runs on WebAssembly inside your browser, so the text and the generated audio never leave your device.
Where does the model come from?
The ONNX graphs are fetched from Hugging Face at run time, not from 73hi. If the Hub is unreachable from your network, you can set a mirror URL or pick model files you already downloaded.
How large is the download?
The speech model and the audio codec together are roughly 760 MB. Your browser caches them, so later visits in the same browser start much faster.
Does it need a GPU?
No. MOSS-TTS-Nano is a 0.1B parameter model designed for CPU inference and runs on the WebAssembly backend.