Consider connecting `KittenTTS` model

[xamgore]

It's small (25 to 80 MB), and sounds better than many default voices:
https://github.com/KittenML/KittenTTS

[ken107]

Thank you, we are indeed waiting on them, and PocketTTS as well.

[luc42ei]

I integrated Supertonic as an offline TTS engine in my fork: https://github.com/luc42ei/read-aloud

It runs entirely in-browser via ONNX Runtime Web (WASM, Web Worker) — no external service. Models (~250 MB) are downloaded on demand from HuggingFace and cached persistently in the browser Cache API. 10 voices (5F + 5M), supports en/ko/es/pt/fr. Quality is noticeably better than espeak.

Install trigger is a dropdown option; progress shown inline in the settings page.

[xamgore]

I'm sure you did a great job, the problem is that nowadays trust is everything and malware can get easily. Do you have plans on publishing it on addongs.mozilla.org?

[ken107]

I integrated Supertonic as an offline TTS engine in my fork: https://github.com/luc42ei/read-aloud

It runs entirely in-browser via ONNX Runtime Web (WASM, Web Worker) — no external service. Models (~250 MB) are downloaded on demand from HuggingFace and cached persistently in the browser Cache API. 10 voices (5F + 5M), supports en/ko/es/pt/fr. Quality is noticeably better than espeak.

Install trigger is a dropdown option; progress shown inline in the settings page.

We have support for Supertonic along with Piper in the Experimental Voices section. Works similarly to your implementation.

[luc42ei]

The implementations are quite different. In the upstream extension, clicking a Supertonic voice opens supertonic.ttstool.com and asks you to install a separate Chrome extension (or use browser cache that may be auto-deleted). My fork runs the ONNX model entirely locally — downloaded once from HuggingFace into the browser Cache API (persistent), synthesized in a Web Worker via ONNX Runtime WASM. No external service, no separate extension required.