nano-edge-tts

About

A 1.75M-parameter replacement decoder for the Kokoro-82M text-to-speech model. It leaves Kokoro's text side and voices unchanged and swaps the 53.3M-parameter waveform decoder for two small networks that set per-frequency gains on a pulse train and on noise, so no neural network runs at audio rate. The model card reports the decoder running 33 times faster on one CPU thread, with word error rate and UTMOS scores close to Kokoro's.