LoudKit: local TTS with voice cloning, 10 languages, and SDKs for Python, Swift, Go, Rust and TypeScript

hey guys, I've been working on a reading app for several months now and had problems with getting good quality TTS, the options were kokoro, kitten, pocket but all of them even though they were sounding natural had some problems when listening longer. Last month I took upon myself to try to get a model that is running on edge (I had an iphone 14 pro as a testbed) and got to what I now packaged as loudkit. It supports 10 languages now, voice cloning, is quite small and fast enough with quality similar to Chatterbox to my ears which was the base model I started optimization from. What is not part of this release is the emotion axis with tags, something I am working on right now. Code and model weights are Apache 2.0. I also ported it (with CC help ofc) to a few languages, because in the past I lost like a week for parsing one TTS tokenizer from python to swift and would lose my mind when I'd get crashes and memory leaks. Here the contract was to get the same speech tokens in all adapters, so it doesn't sound nice in python but sucks in typescript. Audio samples can differ slightly between backends, and file metadata like timestamps can differ too. There are two variants loudr-1 and loudr-1-turbo. basically turbo was done by attaching another head to the most time consuming component of the pipeline and training it so it predicts two audio tokens at once. It worked quite well but sometimes I can still hear the tts artifacts, so YMMV. Voice cloning works quite well but I found the best is to give it around 10 seconds of recording, and if there are long pauses or noise in the background the cloned voice is suboptimal. All included voices come from consented donations or CC0 / CC-BY recordings, with sources documented. repo: https://github.com/loudreader/loudkit docs: https://loudreader.github.io/loudkit/ hf: https://huggingface.co/loudreader/loudr-1 & https://huggingface.co/loudreader/loudr-1-turbo I've seen that the localTTS that can be connected to agents like hermes or openclaw still has issues with quality and thought why not opensource it. Ah, for quality of other voices than english I'm not sure. I sent snippets around and got positive feedback but can't vouch for these. Feel free to check it out, hope you like it.