justdataplz/pocket-tts-web

Pocket TTS for the browser

Kyutai's Pocket TTS and Myned AI's Greek Pocket TTS, converted to run in a web browser with onnxruntime-web on WebAssembly (one thread). Used by the read-aloud voices of everkept.net and of the art museum at artmuseum.artfrompixels.com.

One folder per language (greek, english, french, german, spanish), each about 116 MB:

File	What it is
bundle.json	sizes, text rules and the list of files
text_conditioner.onnx, flow_lm_main.onnx, flow_lm_flow.onnx, mimi_decoder.onnx	the four networks, int8
tokenizer.json	the language's SentencePiece (unigram) model as plain data
mimi_init.bin	the decoder's starting state
voices/<voice>.bin	a voice: the flow LM's attention cache after hearing the speaker, float16

The flow LM uses the KV-delta contract of lookbe/pocket-tts-onnx-export (MIT) with a 400-row attention cache (16-bit): it returns only the rows it adds. The browser loop follows KevinAHM/pocket-tts-web (Apache 2.0) and Kyutai's pocket_tts.

Voices

Language	Voices	Source
Greek	eleni, athina, katerina, spyros, kostas, nikos	Myned AI, from Mozilla Common Voice (CC0)
English	alba, jane, george, michael	Kyutai: Alba Mackenna (CC BY 4.0); VCTK corpus (CC BY 4.0)
French	estelle	Kyutai (CC0 recording)
German	juergen	Kyutai
Spanish	lola	Kyutai, from Mozilla Common Voice (CC0)

Licence and use

CC BY 4.0, as the models it comes from: Pocket TTS by Kyutai (Manu Orsini, Simon Rouard, Gabriel De Marmiesse, Václav Volhejn, Neil Zeghidour, Alexandre Défossez); the Greek model by Myned AI. Converted for the browser by justdataplease.

As Kyutai asks: do not use it to imitate or clone anyone's voice without their explicit and lawful consent, nor to present what it says as a genuine recording of a real person or event, and comply with all applicable laws.