I wanted a portable CPU implementation of kokoro for a project I am working on for a voice interface and found the current GGML/GGUF models were sometimes slower than python torch code. I love Kokoro for it's consistency and desired an easier way to integrate it into my projects. Focusing on a non-quantized SIMD yielded excellent performance for my targets. Theoretically portable to NEON and other accelerators though I have not tried. Hopefully useful for those of you looking for a less complex deployment of a TTS function that includes a G2P.

On the side I am exploring IPA more in depth and is where the g2p lexicon comes from. Espeak-ng involves tradeoffs on lexicon, emphasis, and frankly I just wanted to learn more about the subject. This kokoro work is largely in support of that exploration, but I figure it could be useful to somebody.