Developer open-sources Ninada: 48 browser-ready TTS models

On November 5, 2026, developer h3manth announced on the Reddit community LocalLLaMA that Ninada is now open source: 48 ONNX TTS checkpoints meant to run inside the browser, released under MIT for the code and CC-BY-4.0 for the weights. The checkpoints cover Classical Sanskrit, English, and 10 Indic languages, with each model ranging from 21.7 MB to 27.9 MB. A companion in-browser studio is live at https://h3manth.com/ai/ninada/, requiring no sign-up and running inference entirely on the local machine; the author reports roughly 538ms per synthesis on a phone via WASM. In the post, the author describes building an in-browser TTS model for Indic languages, including Classical Sanskrit, as a longtime personal goal, and explains that most existing Indic TTS setups either depend on cloud APIs, ship multi-gigabyte unified checkpoints whose accents bleed across scripts, or rely on 16 kHz Meta MMS VITS (1.5 GB in total). A follow-up write-up covers why the project was built and what broke on real phones. Based on the available material, the weights, the code, and the online studio have all been released, and no further developments have been reported since the launch post.