I built a local TTS add-on using an 82M parameter neural model – Runs on potatos

Wanted to share something I’ve been working on: a Firefox add-on that does neural-quality text-to-speech entirely offline using a locally hosted model.

No cloud. No API keys. No telemetry. Just you and a ~82M parameter model running in a tiny Flask server.

It uses the [Kokoro TTS](https://huggingface.co/spaces/hexgrad/Kokoro-TTS) model and supports multiple voices. Works on Linux, macOS, and Windows but not tested

Tested on a 2013 Xeon E3-1265L and it still handled multiple jobs at once with barely any lag.

Requires Python 3.8+, pip, and a one-time model download. There’s a .bat startup option for Windows users (un tested), and a simple script. Full setup guide is on GitHub.

GitHub repo: https://github.com/pinguy/kokoro-tts-addon

Would love some feedback on this please.

Hear what one of the voice examples sound like: https://www.youtube.com/watch?v=XKCsIzzzJLQ

To see how fast it is and the specs it is running on: https://www.youtube.com/watch?v=6AVZFwWllgU

---

| Feature | Preview | | ---------------------------------------------------------------- | --------------------------------------------------------------------------------------- | | *Popup UI*: Select text, click, and this pops up. | [](https://i.imgur.com/zXvETFV.png) | | *Playback in Action*: After clicking "Generate Speech" | [](https://i.imgur.com/STeXJ78.png) | | *System Notifications*: Get notified when playback starts | (not pictured) | | *Settings Panel*: Server toggle, configuration options | [](https://i.imgur.com/wNOgrnZ.png) | | *Voice List*: Browse the models available | [](https://i.imgur.com/3fTutUR.png) | | *Accents Supported*: American English, British English, Spanish, French, Italian, Portuguese (BR), Hindi, Japanese, Mandarin Chines | [](https://i.imgur.com/lc7qgYN.png) |

---